DEVELOPING GAN’S BASED ETHIOPIAN TRADITIONAL

DRESS FASHION DESIGN IMAGE GENERATION MODEL

DEVELOPING GAN’S BASED ETHIOPIAN TRADITIONAL DRESS
Mikiyas Zerfu FASHION DESIGN IMAGE GENERATION MODEL
AASTU 2021
 A MASTER’S THESIS

BY

MIKIYAS ZERFU TESFA

DEPARTMENT OF SOFTWARE ENGINEERING

ADDIS ABABA SCIENCE AND TECHNOLOGY UNIVERSITY

FEBRUARY 2021
DEVELOPING GAN’S BASED ETHIOPIAN TRADITIONAL DRESS
FASHION DESIGN IMAGE GENERATION MODEL

By

MIKIYAS ZERFU

A Thesis Submitted as a Partial Fulfillment for the Degree of Master of Science in software

Engineering

To

DEPARTMENT OF SOFTWARE ENGINEERING

ADDIS ABABA SCIENCE AND TECHNOLOGY UNIVERSITY

FEBRUARY 2021
 Certification
This is to certify that the thesis prepared by Mr. Mikiyas Zerfu Tesfa entitled “Developing

Gan’s Based Ethiopian Traditional Dress Fashion Design Image Generation Model”

submitted as a partial fulfillment for the Degree of Master of Science complies with the

regulations of the University and meets the accepted standards with respect to originality,

content and quality.

Singed by Examining Board:

External Examiner: Signature, Date:

Signature, Date:
Internal Examiner:

Signature, Date:
Chairperson:

DGC Chairperson: Signature, Date:

College Dean/Associate Dean for GP: Signature, Date:

 Declaration
I hereby declare that this thesis entitled “Developing Gan’s Based Ethiopian Traditional

Dress Fashion Design Image Generation Model” was prepare by me, with the guidance of

my advisor. The work contained herein is my own except where explicitly stated otherwise

in the text, and that this work has not been submitted, in whole or in part, for any other

degree or professional qualification.

Author: Signature, Date:

Witnessed by:

Name of student advisor: Signature, Date:

Name of student co-advisor: Signature, Date:

Abstract
Technology-driven breakthroughs in the fashion industry, such as artificial intelligence (AI) or
machine learning, are revolutionizing every component of this forward-thinking business area in
the twenty-first century. Deep learning, algorithms, and big data have a broad range of
applications in the fashion industry, despite the fast advancement; this paper presents a
comprehensive review on Application of AI in fashion-related works and covering Fashion
Image Generation, Fashion Recommendations, Fashion e-commerce and Fashion Virtual tray-
on. The purpose of this research is to explore and assess the use of Generative Adversarial
Networks in the creation of new attractive outfits, as well as to deal with the development of
Ethiopian Habesha fashion clothing images. Three components make up the proposed deep
learning model: (1) Creating a unique fashion image from ready-to-wear Habesha garments. (2)
Creating a creative design picture based on fashion dress patterns and designs (3) finally, using
the preceding two outcomes, image-to-image translation and produce novel fashion dress
design. This study, together with the Framework, explains how to utilize GAN to fulfill the
process of fashion design. Also, offer a mechanism for obtaining and cleaning picture data
automatically, as well as a dataset including over 13,000 images of Ethiopian clothing and over
500 designs and illustrations. Comprehensive study has proven that the selected model can be
highly useful in the apparel industry. There is additional information about the lessons learned,
as well as a discussion of existing constraints and potential future possibilities.

Keywords: Generative Adversarial Network, Deep Learning. AI, Fashion Design.

i
 Acknowledgement
First and foremost, I would like to praise and thank God, the almighty, who has granted
countless blessing, knowledge, and opportunity to the writer, so that I have been finally able to
accomplish the thesis.

Apart from the efforts of me, the success of this thesis depends largely on the encouragement and
guidelines of many others. I take this opportunity to express my gratitude to the people who have
been instrumental in the successful completion of this thesis. The gratitude goes to all those who
have assisted me in the process of completing this thesis. It would be impossible to list all names
but several people deserve my sincere and special thanks.

My special deepest gratitude and heartfelt thanks go to my advisor and computer vision special
interest group Leader Dr Tulu Tilahun (Ph.D.) and Director of center of Excellence of AI and
Robotics for his persistent guidance and support, and supervision from proposal to the
completion of this thesis. Without his encouragement and guidance, this thesis would not have
materialized.

My Second special thanks go to Mr. Ashenafi Kebede(MSc.), who is a postgraduate program

coordinator, Dr. Hussein Seid (Ph.D.), and Dr. Kula Kekeba (Ph.D.) for their voluntary
evaluation of my work and for the valuable inputs to be included in this thesis and generously
spent precious time in giving the guidance, encouragement, suggestion, and comments for this
thesis.

I am also indebted to Mr. Ezedin Wangori as provider and subscriber for my research
instruments and tools.

Finally, my thanks go to my Collogues for their effort during dataset collection and annotation. I
also thanks Fast track master’s program students, friends, and parents for their relentless support
and advice in my life. May Almighty God always Keep safe and bless them through their life.

It would be difficult to find adequate words to convey how much I owes the people. Lots of love
and thank to all of you.

May God bless us.

ii
Mr. Mikiyas Zerfu
February 2022

iii
Table of Contents
Abstract...........................................................................................................................................................i
Acknowledgement.........................................................................................................................................ii
Lists of Tables...............................................................................................................................................vi
List of Figures..............................................................................................................................................vii
Abbreviations and Acronyms........................................................................................................................1
Chapter 1. Introduction..........................................................................................................................2
1.1. Background....................................................................................................................................2
1.2. SWOT Analysis.............................................................................................................................6
1.3. Motivation......................................................................................................................................7
1.4. Statement of the problem...............................................................................................................8
1.5. Objectives....................................................................................................................................10
1.5.1. General Objective................................................................................................................10
1.5.2. Specific Objective................................................................................................................10
1.6. Scope and Limitation of the Study...............................................................................................10
1.7. Significance of the study..............................................................................................................11
1.8. Organization of the Thesis...........................................................................................................12
Chapter 2. Literature Review..............................................................................................................13
2.1. Introduction..................................................................................................................................13
2.2. Background Concepts..................................................................................................................14
2.2.1. Definition of terms...............................................................................................................14
2.2.2. Traditional Fashion Design and Supply chain.....................................................................15
2.3. Fundamentals of Artificial Intelligence.......................................................................................18
2.3.1. Machine Learning................................................................................................................19
2.3.2. Deep learning.......................................................................................................................22
2.4. Generative Adversarial Network.................................................................................................26
2.4.1. Common GAN Architecture................................................................................................27
2.4.2. Conditional Generative Adversarial Nets............................................................................28
2.4.3. Deep Convolutional Generative Adversarial Networks......................................................29
2.4.4. Cycle-Consistent Adversarial Networks..............................................................................29
2.4.5. A Style-Based Generator Architecture for GANs...............................................................30
2.4.6. Progressive Growing of GANs............................................................................................31
2.5. Overview of deep learning applications for Fashion Industry.....................................................32
2.5.1. Deep Learning Fashion Image Generation..........................................................................33

iv
 2.5.2. Deep Learning for Fashion Recommendations....................................................................35
2.5.3. AI for Fashion e-commerce.................................................................................................37
2.5.4. AI for Fashion Virtual tray-on.............................................................................................38
2.6. Generative Adversarial Networks for Fashion Industry..............................................................41
2.7. Related Works..............................................................................................................................43
Chapter 3. Methodology.......................................................................................................................48
3.1. Overview of Methodology...........................................................................................................48
3.2. Concept Development..................................................................................................................49
3.3. Dataset Preparation and Preprocessing........................................................................................50
3.4. Loss Functions in GANs..............................................................................................................52
3.4.1. Cross-entropy loss................................................................................................................52
3.4.2. Binary Crass-Entropy...........................................................................................................52
3.4.3. Adversarial Loss..................................................................................................................52
3.4.4. Losses from Additional Generators.....................................................................................53
3.5. Research Tools.............................................................................................................................54
3.5.1. Software Tools.....................................................................................................................54
3.5.2. Hardware Tools....................................................................................................................55
3.6. Evaluation Metrics.......................................................................................................................56
3.6.1. Fréchet Inception Distance (FID)........................................................................................56
3.6.2. Inception Score....................................................................................................................57
3.6.3. Nearest Neighbors................................................................................................................58
Chapter 4. Design And Experiment....................................................................................................59
4.1. Proposed Model and Model Description.....................................................................................59
4.2. Dataset Preprocessing..................................................................................................................61
4.3. Experimental Setup and Implementation.....................................................................................62
4.4. Implementation of Model Evaluation..........................................................................................65
Chapter 5. Results and Discussion.......................................................................................................66
5.1. Gathering of Image Data..............................................................................................................66
5.2. Generated Samples with different Iterations................................................................................67
5.3. Generated Image from StyleGAN...............................................................................................69
5.3.1. Failed examples from StyleGAN.........................................................................................70
5.4. Generated Image from StyleGAN2 with adaptive discriminator augmentation (ADA).............71
5.5. Evaluation Results.......................................................................................................................72
5.5.1. Quantitative results..............................................................................................................72
5.5.2. Nearest Neighbors results....................................................................................................72

v
 5.6. Discussion....................................................................................................................................75
Chapter 6. Conclusions and Recommendations.................................................................................77
6.1. Conclusions..................................................................................................................................77
6.2. Limitations and Recommendations..............................................................................................77
References...................................................................................................................................................79
Appendices..................................................................................................................................................88
Appendix A: Scrape Parameters..............................................................................................................88
Appendix B: Participant Interview questions..........................................................................................89
Appendix C: Original Sample Dataset.....................................................................................................91
Appendix D: Sample Results from the Training of StyleGAN...............................................................92
Appendix E: Sample Results from the Training of StyleGAN2AdaIn....................................................93

vi
Lists of Tables
Table 1. SWOT Analysis for Fashion Industry and Technologies-------------------------------------6

Table 2. Adapted version of the scheme of Norvig and Russell--------------------------------------18

Table 3. Different Machine Learning Algorithms-------------------------------------------------------20

Table 4. Artificial Intelligence Application in Fashion Industry--------------------------------------32

Table 5. Related works in detail---------------------------------------------------------------------------46

Table 6: Dataset Description-------------------------------------------------------------------------------51

Table 7: Generated Images Using StyleGAN Model---------------------------------------------------69

Table 8: failure examples-----------------------------------------------------------------------------------70

Table 9: Generated Image from StyleGAN2 with adaptive discriminator augmentation (ADA)-71

Table 10 : List of Networks snapshot with their respective FID result of StyleGAN--------------72

Table 11: List of Networks snapshot with their respective FID of StyleGAN2 with adaptive

discriminator augmentation (ADA)-----------------------------------------------------------------------72

vii
List of Figures
Figure 1.Ethiopian Traditional Fashion Cloth maker (shemane) [33]--------------------------------15

Figure 2. Traditional apparel supply chain. [35]--------------------------------------------------------17

Figure 3. Artificial Neural Network architecture--------------------------------------------------------21

Figure 4 How deep learning can be a set of machines learning and the way ML is a set of AI

[47].------------------------------------------------------------------------------------------------------------22

Figure 5. Deep Learning Architecture--------------------------------------------------------------------22

Figure 6. The architecture of a CNN that performs classification of Animals----------------------23

Figure 7. Recurrent Neural Network with two hidden layers------------------------------------------24

Figure 8. LSTM containing four interacting layers example.-----------------------------------------25

Figure 9. Generative Adversarial Networks Architecture----------------------------------------------27

Figure 10. Simple Cycle Gan Architecture---------------------------------------------------------------29

Figure 11. StyleGAN Architecture------------------------------------------------------------------------30

Figure 12. Progressive Growing of Gans Sample-------------------------------------------------------31

Figure 13. Virtual tray on in store.------------------------------------------------------------------------39

Figure 14. Concept Development Process---------------------------------------------------------------49

Figure 15. Web Scrapping Process------------------------------------------------------------------------50

Figure 16: The Proposed Model framework-------------------------------------------------------------60

Figure 17: Dataset Random capturing's-------------------------------------------------------------------64

Figure 18: Dataset tfRecords result-----------------------------------------------------------------------64

Figure 19: List of Images based on Training Iterations.------------------------------------------------68

Figure 20: Nearest Neighbors assessment from random people.--------------------------------------73

viii
Abbreviations and Acronyms
Acronyms Definition
AI Artificial intelligence
AttGAN Attribute Generative adversarial networks
CADD computer aided design and drafting
CGAN Convolutional Generative Adversarial Network
CNN Convolutional Neural Networks
CycleGAN Cycle Generative Adversarial Network
DCGAN Deep Convolutional Generative adversarial nets
FID Fréchet Inception distance
FN False Negative
FP False Positive
GAN Generative adversarial networks
Gb Giga byte
GDP Gross domestic product
GPU Graphics processing unit
IS Inception score
ML Machine learning
NLP Natural language processing
Open-CV Open Source Computer Vision Library
ProGAN Progressive grows Generative adversarial networks
STYLEGAN Style Generative Adversarial Network
TN True Negative
TP True Positive

1
Chapter 1. Introduction
1.1.Background
Clothing is one of the images and manifestation of human civilization. The clothing trends and
designs represent man's search for beauty from ancient times to the present. As time progresses,
human clothes change in various times, but beauty and creativity are constantly pursued. Fashion
design is an art frame that combines practicality and creativity. The fashion industry could be a
worldwide commerce of 1.3 trillion dollars, which utilizes more than 300 million individuals
around the world [1] and represents a critical economic drive and a considerable force of global
GDP. According to the showcase analysis provided by a few of the most universal business
consultancy companies (McKinsey, Deloitte Bunch, BCG), fashion companies today are working
in a really dynamic competitive environment, dominated by sudden changes and expanding
vulnerability. The fashion industry has always been at the forefront of technological
improvement, whether it is through the sewing machine or the e-commerce market [1].

Innovation during this fashion and attire sector does not solely rely upon product innovation;
however the new digital technologies applied to style and production processes are more
changing into crucial. Rising technologies features a vital potential to alter whole production
processes and business models for the textile and clothing industries with an intense work force.
Technology-driven developments in the fashion industry, such as artificial intelligence (AI) or
machine learning, are transforming every area of this forward-thinking business domain in the
twenty-first century.

Ethiopia has strengths throughout the whole value chain from fiber, yarn, to clothing across the
complete worth chain from fiber, yarn, material to apparel. The fashion, textile and clothing
industries are it offers a wide range of categories from traditional handmade items, handicrafts,
wool and silk to the organized textile industry. Ethiopia's textile and clothing sector has
enormous potential, having expanded by 51% on average over the previous 5-6 years [2] . With a
key target, Foreign Direct Investments (FDI) the government is driven to open up their textile-
and attire market to mid-in massive scaled foreign investors. Ethiopia's economy is largely
dependent on agriculture, although the government has worked hard over the last five years to
diversify exports, with a focus on critical industries such as light manufacturing, textile

2
manufacturing, and garment manufacturing. In the previous 5–6 years, the textile and clothing
sector in Ethiopia has developed by an average of 51%, with about 65 international investment
projects authorizing foreign investors. In order to generate future employment and to achieve its
objective of improving profits from foreign currency, the Ethiopian Government has led the
Fashion industry as a key priority this is often supported by the creation of varied trade policies
and incentives to draw in foreign direct investors. In order to generate future employment and to
achieve its objective of improving profits from foreign currency, the Ethiopian Government has
led the industry as a key priority. This is supported by the creation of various industry policies
and incentives to attract foreign direct investors. At present, there are over 130 medium- and
large-scale enterprises in the textile and clothing industry, 37 of which are externally owned. In
2014, the sector's export size reached around USD 113 million and 37,000 employees were
engaged. Ethiopian textile and clothing industry benefits from inherent advantages in raw
material and simple availability of trained labor, as well as a rich basis of diverse traditional
design, color, and embroidery combinations [3].

The traditional Ethiopian fashion has always been a need and a demand for many people in
Ethiopia and around the world. Over the last decades, there has been a distinct traditional
appareling need in our country to change significantly in order to meet the needs of society [4].
Interestingly, traditional fashion has been, of course, very competitive across our country on the
local market. Though traditional fashion design method was the good approach to produce the
right fashion items, there were several problems regarding manual design process like time
consuming, subjectivity and expensive work force. Modern approaches to fashion design vary
from the use of the software tools like CADD and Photoshop to new approaches for artificial
intelligence and machine learning. We have continuously utilized our technological framework
to push ourselves past our natural restrictions. Our machines are not continuously working
against us, but rather working as an expansion of our minds. Additionally, we seem utilize our
machines to grow on our creativity and thrust the boundaries of art.

Today we already continue to see revolutions that would have transformative cultural and social
consequences in fields such as AI and Machine learning methods to artistic activities such as
image and music synthesis [5], as well as style transition have increased in popularity. Artwork
is increasingly being created by machines through algorithms with little or no input from

3
humans. AI art has emerged in the later part of the 2010s as a more recent trend in history,
drawing on generative and algorithmic approaches to produce works partially or entirely
produced by some kind of machine learning [6].

The utilize of AI within the fashion industry of 2020 has ended up obvious since in 44% of the
design retailers are nowadays confronting bankruptcy. As a result of this, worldwide investing on
AI advances by the fashion & retail industry is anticipated to reach $7.3 billion each year by the
year 2022. Supported by the simple accessibility of big data, client personalization, and other
services in fashion, companies are basically not attainable without the use of AI in fashion.
According to McKinsey [1], the driving 20% of the worldwide fashion brands are producing
144% of the industry profits. This means that for any design brand, it should be within the best
20% to be a beneficial business. It is high time Fashion and Textile education should, without
doubt incorporate latest and emerging technologies to develop students and youngsters to be
suitable to handle the future of the industry. Driven by this need, fashion brands are investing in
AI and ML innovations to stay important in a profoundly competitive commercial center. AI has
started to enter the field of art and fashion by a step-by-step trial and exploration while the fast
advancement and the broad use in areas such as computer vision and natural language processing
(NLP). In labor-intensive textile manufacturing, the implementation of AI was not increasingly
accepted. Nevertheless, AI's broader implementations in the fashion industry are primarily due to
its market competition and the goal of low production costs from Attire designing,
Manufacturing process, to Virtual merchandising [7].

AI has a wide variety of applications in the fashion industry through the use of Deep learning,
algorithms, and big data to increase human and market capacities. Such Applications are design
and product creation [8] that help new product creation by analyzing design components such as
color, cloth, shapes, and cut. As a designer [9], AI creates innovative fashion designs along with
sewing templates by using a powerful algorithm that analyzes past designs and future trends. AI
in textiles can be used in clothing production by automating cloth quality management, pattern
checks, color matching, and defect identification. From the point of view of purchase and
merchandising [10], AI used in predictive analysis by automatic product labeling and pattern
predicting, personal styling platforms, visual merchandising platforms, and performance
measurement. Online product and visual resemblance recommendations complete the look

4
recommendations, and customized recommendations per country, per user section, and per
individual customer are all examples of AI in e-commerce [11]. AI can be used in automated
product tagging, natural language search, visual search and chat-bot support for customer care
and client serving, branding strategies, and marketing strategy [10] [12] [13] [14].

The image generation task becomes simple after the introduction of the generative adversarial
network [15]. There are two types of image generation technology: generative parametric and
adversarial approaches [16]. Several studies have shown that adversarial approaches, particularly
the generative adversarial network, have clear advantages in image generation (GAN). GAN is
type of machine learning in which two adversarial models are trained at the same time: a
generator ("the designer") that learns to generate images that appear real, and a discriminator
("the design critic") that learns to discern real images from fakes. During preparation, the
generator improves its ability to create realistic images, while the discriminator improves its
ability to detect fakes. This is used dynamically to produce computer-generated images and
motion that looks plausible (and possibly aesthetically pleasing) to the consumer [17].

The creation of different advanced GANs models like Conditional GAN (CGAN), deep
convolutional GAN (DCGAN) algorithm [18], Progressive Growing of GANs (ProGAN) [19],
Style-based Generative Adversarial Network (Style Gan) [20], and CycleGAN [21] is a
significant step forward in the development of the GAN model, as it establishes a robust training
framework that makes creativity implementation easier.

Inspired by such advantages and from some researches has been done in the past that has shown
the importance of the generative adversarial network in the creation of fashion products and
image-to-image translation [22], this research plan is to deal with the generation of Ethiopian
Habesha dress images in style, design, color and other different conditionals with considering
Patterns. Also, this study planning to analyze and evaluate application of GANs for generating
new fashionable clothes.

Aside from the main objective of the research, this research includes perspectives on fashion and
textile education, textile and clothing industries, challenges and possibilities, innovation in
fashion and textiles, entrepreneurial development for fashion and textiles, fashion, digital
technology and sustainability, skill development, and fashion marketing, communications and
the development of clothing and clothing industry.

5
 1.2.SWOT Analysis
The SWOT analysis is a strategic planning approach used to evaluate strengths, weaknesses,
opportunities and threats in connection with the competitive or project planning of a person or
organization. An examination of strengths, weaknesses, opportunities and threats is one of the
key phases in the planning process. Prior deficiencies and risks are identified and elements,
which might have detrimental consequences outside the scope of the agriculture sector, can be
found in appropriate methods to internal improvements and justified. Recognition of strengths
and possibilities allows us to achieve as much as possible the aims and goals established by the
internal and external contexts.

The prospects for new Fashion enterprises are investigated with this SWAT analysis report. The
study gives an outline of the potential and limitations of the industry to evaluate the prospects of
investing or starting enterprises, or of cooperating with Ethiopian producers of textiles and
clothing. Supported by this short review however, detail analysis the listing of strengths and
weaknesses provides and overview of the present standing of Ethiopian textile and garment
industry [2].

Table 1. SWOT Analysis for Fashion Industry and Technologies

Strengths (S) Weaknesses (W)

1. Availability of abundant labor at low

cost
1. Low production of local cotton.
2. Strong financial support government
2. Lack of diversification in products, relative low
3. Import duty free equipment, machinery
quality of the products.
and on spare parts and / or any goods
3. The inability to compete on the international
needed to realize production.
market due to price pressure.
4. Trained fashion technologist come
4. No manmade fibers – only through imports
from various Universities- good
5. Bureaucracy
education
6. Lack of merchandisers
5. Industry Parks are very cheap to settle
7. Lack of IT and CAD/ CAM in production.
6. Strong presence of foreign technology
experts.

6
 Opportunities (O) Threats (T)

1. Low costs of commodities like water

and electricity.
2. Constructing strong forward and
1. Political instability and unsafe nation
backward linkages and attracting new
2. Lack of IT and CAD/ CAM in production
prospects.
3. Non-availability of spare parts for maintenance of
3. Speed to market by using airfreight at
Machine Park.
competitive costs
4. Manual pattern and marker making
4. Ethiopian government is to make a
transformation from an agricultural to a
manufacturing economy

1.3.Motivation
In recent years Deep Learning has received a lot of interest, attaining performance that can able
to exceed human performance, obtaining astounding outcomes for specific jobs. Furthermore,
from the Deep Learning area, a branch has gained special notice, Generative Adversarial
Networks. First proposed by Ian Goodfellow in 2014, this sort of training has been utilized in
many areas such as voice improvement, image synthesis, attention prediction, anomaly detection
and Even in art production [15]. Therefore, it is fascinating to examine image generation
algorithms, notably GAN-based approaches, and also use certain sophisticated techniques on
fashion datasets to compare their performance and maybe acquire some insights in this area.

In addition, being an Ethiopian, I understood what the Ethiopian design truly was: it was largely
custom-made. Many times, Ethiopian fashion was simply for events, whether its weddings or
festivals or going to church. There were no other occasions you would see traditional Ethiopian
clothes. Nevertheless, in other regions of Africa, people would wear their traditional attire even
in the office. This got me excited about promoting the sector because I wanted the designers to
start creating ready-to-wear clothing. It was so that people might start being proud to possess
Ethiopian clothing, and instead of wearing them simply for occasions, they would wear
Ethiopian-made clothes day-to-day. It is possible and can be assisted with artificial intelligence.

7
 1.4.Statement of the problem
This study's problem statements can be divide into a few categories. The former is from a digital
fashion design perspective, the next is focused on generative adversarial strategies and
improving processes on architecture that are used to improve image generation efficiency and
image-to-image translation. Moreover, the last one is how to add many conditional inputs to
maintain user engagement, as fashion is a difficult task in and of itself [4].

The big issue in the fashion sector is the manual design process. Even though we have this
sentimental notion of a fashion designer designing his pieces by hand, this will be impractical
because of the rapid speed of fashion cycles. In addition, the design process is usually not
innovative, particularly in the quick fashion industry, but rather focused on current prototypes of
a brand being mixed with new trends or merely updated in cut and color. The fashion industry's
design production methods are expensive for human power and time inefficient. Fashion
designers have to continue manually analyzing trends, collecting data, creating mood boards, and
other procedures [23]. According to previous studies due to the subjectivity of markings and
semantic sophistication of features for the vast development of fashions [24], it is difficult in
realistic design and production processes to satisfy the users' interests. The major challenge is
made even harder to take the user's personalized styles into account, which are generally
dependent on the designer's interest. In the past, computer-aided design and drawing [25] were
introduced to deal with such an issue to ease the design process. It did not take long to use
CADD fashion software because of the large manufacturing of pieces, while manual work for the
design is still required [26]. The inadequate division of capital, not concentrated on the business
customers, arbitrary architecture and over-production, as well as big designers, have been caused
by the inefficiency of the design method. For one set, designers need 1700 hours and color
concepts just 400 hours. Since the statistics and pattern analyses is not exact, as well as
erroneous management choices, the loss is enormous, billions of garments never sold and tones
of contaminating waste are collected together. In addition, there is huge subjectivity and being
dependent on designer interests are issues from fashion design perspective [23].

The second concerns is on computer creativity. Creative worlds play a major part in today's
culture with a significant economic and cultural effect. A great deal of work is made in creative
fields including fashion, art and design, distribution and analytics of creative content [27]. Study

8
on the application of computer vision algorithms for different areas like artistic fields, including
production, analysis and visual information processing have become thematic research areas in
recent years. Creativity of computers and deep learning have become increasingly popular. This
ever-increasing interest is most evident in Computer Vision for Fashion, including new fashion
images generation, e-commerce, virtual try-on and personalized recommendation, etc.

Present methods in actual implementations of deep learning for fashion design are also far from
being implemented. For example, when given a user-taken photo, the quality and reliability of
retrieving clothes in various commercial photos still has a lot of space for improvement. There
are also several issues to be resolved with generating new fashion design images, because the
sophistication of the architecture is too high [4], the image produced is a poor quality, the visual
artifact and the accuracy are blurred. Some of those problems are will be solved by the proposed
model and the remaining questions will be open for future Investigation. Incorporating
generative models into reinforcement learning in a variety of ways allows machine learning to
interact with multi-modal outputs. Subjectively, GANs are thought to provide better samples
than other approaches.

In this study, a Generative adversarial network-based Fashion design image generation method
are proposed that will generate a novel Ethiopian Habesha dress with new design and pattern,
which also, can be blend with African dress patterns. Moreover, minimize visual artifacts and
inconsistency issues with generated results attributes such as form, patterns, texture, color and
size. This research also performs the task of collecting images of fashion sketches and dress
images from various sources and make ready them for generation of adversarial workouts after
several pre-processing.

To this end, this research will try to give an answer the following research question.

 How can generative design, be incorporated in the fashion design process?

 How to improve image-to-image translation for fashion-design image generation?
 How the performance can be improved by devising new or combined learning
frameworks and/or network architectures and/or training data?

9
 1.5.Objectives
1.5.1. General Objective
The main objective of this thesis is to Design and develop a generative adversarial network-
based model for generating a novel interactive fashion design image with different conditions
including pattern that can give aid to the fashion designers.
1.5.2. Specific Objective
In the research proposed, the following are specific objectives to meet the General objective:

 Investigate the domain using comprehensive review.

 Collect data from different source and prepare dataset.
 Preprocessing the collected data.
 Create a GANs based model for generating novel fashion design and fashion images.
 Train the created model.
 Evaluate the performance of the developed model based on useful qualitative and
quantitative measures; i.e. Nearest Neighbors, Fréchet Inception distance.

1.6.Scope and Limitation of the Study

The work process started with a fairly broad scope to become acquainted with the area of deep
learning. When adequate information had been acquired, through reading various papers and
study results, the scope was limited to concentrate on the most relevant topics. Since artificial
intelligence and deep learning are fields of study yet to achieve stability, new results are
published regularly and it is impossible to intuitively, develop a judgment on what network
topologies are most suited for particular tasks without first having a broader perspective.

Although the entire scope of this study is focused with Application of deep learning in various
industries. This thesis is mainly focused on investigating the Application of Generative
adversarial architecture and creating an architecture for multi-conditional in an Image generation
for Habesha fashion industry. The scope spans from gathering and preparing datasets, to creating
dynamic fashion design and fashion image generation, to Evaluating the quality of produced
results. Through a critical engagement with the design process, the study aims to raise concerns,
increase awareness, and provide fresh views regarding creative activity, as well as to offer some
framework for future progress and innovation.

10
 1.7.Significance of the study
This research propose a new task in generative adversarial networks fashion studies. Instead of
using text instructions, simulated try-on, or texture transferring to create images or using only
one model of GANs, this research challenge is to create new fashion styles using disentangled
user-assigned attributes and conditionals based on two or more combined model.

This study will have great significance in both academia and industry. The main beneficiaries of
this study can be applied in various domains such as designers which can be used to easily build
designs as a creative method, fashion consumers to explore new fashion that are not worn before,
image-to-image translation and style transfer in clip summarization, image enhancing, and
painting arts. Also, for improve quality and control the synthesis, color, shapes, texture and
patterns and it has a potential to complement and enhance the respective creative process.

A new dataset, HabeshaDressSet, were introduced. Unlike current fashion datasets, where most
photographs depict the user's face or body parts with complicated backgrounds and little
inclusion of Habesha’ n dresses, this dataset will depict the user's face or body parts with a
simple context.

In addition to these, the significance of this thesis is described as follows:

 Propose new GANs model for automating the process of fashion design.
 Collect and preprocess a unique large Ethiopian dress dataset (>15000) that contains
Habesha dress sketch's and Habesha dress images and pull it to public repositories.
 Survey for Applications of Modern AI Technologies how to improve Traditional way of
fashion design.
 Producing numerous unknown fabric fashions using users’ preferences, and painting arts.
Under Style transfer.
 It may be used as a design tool to rapidly produce prototypes and assist design processes
and innovation in the manufacturing industry.
 Advise and encourage businesses to use artificial intelligence in their product design
process and serve as a stepping-stone that helps organizations take the first step into a
new method of creating goods.
 Generation of new Ethiopian does not exist faces that can helpful for cartoons and art.

11
 1.8.Organization of the Thesis
Chapter 1: describes the background, motivation, statement of the problem, objectives, scope,
and limitation, introduction to the methodology, and the contribution of research.

Chapter 2: discusses the background literature and related works regarding fashion. A review
will be presented, where the limitations and performance of state-of-the-art methods for
conditioned image synthesis will be summarized. Besides, several basic models acting as the
building blocks of employed approaches and some machine learning techniques will be briefly
introduced.

Chapter 3: presents the methods from data collection, preprocessing, and annotation techniques
and introduce the created fashion datasets, relevant experiment settings and evaluation metrics
for results and models.

Chapter 4: covers the proposed solution in a detailed flow chart, block diagrams, and pseudo-
code in both training and testing algorithms with the equivalent mathematical expression. And,
discusses the implementation step of the proposed solution, implementation sample source code
of the proposed solution.

Chapter 5: presents the proposed solution results and compares the new result of the proposed
solution with the previous solution and elaborates on the meaning of the results.

Chapter 6: summarizes and concludes the work along with the result and it explores the open
issues in the research

12
Chapter 2. Literature Review
2.1.Introduction
Technology has always been a component of dressmaking since the dawn of time. The
foundations for transforming the sector from a small manufacturing system to one of huge
industry were set thanks to the introduction of machines into the production process. The fashion
business, like many other industries, is benefiting from the digital revolution by increasing its
ability to create and use data that was previously not technically possible. As a result of the
development of new data and information through digital technology, the fashion industry may
be better understood and controlled, decreasing uncertainty and boosting coordination. The huge
use of data, increasing computer power, and the availability of advanced algorithms and critical
analytics are all contributing to a major growth in artificial intelligence application. This trend
will continue in the fashion sector, with artificial intelligence being applied in many aspects of
the business. The most often used applications are in the field of customer service, allowing
businesses to capitalize on the growing trend of client customization by improving customer
experience both online and in shops.

This chapter began with presenting a comprehensive picture of fashion design in both classic and
modern fashion design concepts, as well as the supply chain. Second, an introduction and review
of some basic knowledge of models that are building blocks of approaches in Chapter 3, as well
as an introduction to deep neural networks, including how they can be applied for learning useful
and relevant representations, how they can be trained, and how they can be constructed by
employing modular layers of different categories. Then, go over the principles underlying
generative adversarial networks (GANs), with their many topologies illustrated in depth.
Including Deep Convolutional Generative Adversarial Networks (DCGAN), Conditional
Generative Adversarial Networks (CGAN), Progressive grows Generative adversarial networks
(ProGAN), and Style based Generative Adversarial Networks (StyleGAN). Techniques and
tricks for stabilizing the training process of GANs are also examined, and finally, related works
in generative adversarial are discussed, as well as their limitations.

13
 2.2.Background Concepts
2.2.1. Definition of terms
In consideration of the various components of the supply chain, which contribute to the
production of industrial materials into products and the delivery of products to the end customer,
the textile and clothing sector is one of the world's largest industries. in addition to the
production and subsequent wholesale and retail of fiber, yarns, textiles, fashion for men, women
and kids, home furnishers and textile soft products, the textile and clothing industry supply chain
comprises. The study of the Fashion industry is complex and encompasses a wide variety of
disciplines including aesthetics, design, product development, economics, global business,
marketing, consumer behavior, operations, and management as well as anthropology,
psychology, and sociology.

The following are common terms and their definition in fashion design sector:

Fashion: Although “fashion” is frequently used as a synonym for the current dress style in
clothes, sociologists and other researchers who write about fashion are more likely to adopt a
definition that includes two parts. It is (1) accepted by a large number of individuals and (2)
acceptable for a short length of time. In all civilizations and historical times, fashion does not
exist. It appears to have started in the late middle Ages in Western Europe. The nobles created
trends during the period, and the lower classes followed them (a phenomenon known as the
"trickledown theory of fashion.") [28].

Fashion Design:

Definition 1: Fashion design is a form of art dedicated to the creation of clothing and
other lifestyle accessories [29].

Definition 2: Fashion design is the art of applying clothing and accessories in design,
esthetics, apparel and natural beauty. It is affected by culture and various trends and over time
has altered. A fashion designer makes clothes for consumers, including clothes, suits, pants,
skirts and accessories such as shoes and handbags. He or she can specialize, or work in more
than one of these fields, in clothes, accessories or jewelry creation [30].

Traditional clothes: Traditional clothing may be described as a set of past based clothes,
jewelers, accessories worn by an identity group. While minor color, shape and material changes

14
are recognized throughout time, the assembly seems to be transmitted unmodified in the past
[31].

2.2.2. Traditional Fashion Design and Supply chain

Traditional Fashion Design

Culture industries employ creativity, cultural knowledge to generate socially and culturally
significant products and services. Culture industries employ creativity, cultural knowledge to
generate socially and culturally significant products and services. Also referred to as traditional
cultural expression or expressions of folklore, spiritual and emotional features that characterize a
society or social group. The apparel business offers numerous advantages to daily life all around
the world. Fashion extends beyond simply wear and tears to communicate your personality,
create well-being and encourage creativity. The traditional cottage textile sub-sector has a long
history in Ethiopia. People have been making traditional garments out of cotton given by small-
scale farmers and home spun yarns for a long time. Cloth making was passed down from
generation to generation, with Handlooms being used in both rural and urban regions for
livelihood and financial gain generations [32].

Figure 1.Ethiopian Traditional Fashion Cloth maker (shemane) [33]

15
Ethiopia has cotton-producing agricultural land. Cotton cultivation and clothing manufacture
have a long history in the country. Cotton is one in every of the oldest cultivated fabricator crops
and major money crops in Ethiopia.

In Ethiopia, traditional textiles focused on the status of the nation as a cotton cradle. In Ethiopia,
a traditional and home-generated business is the handloom industry. Weavers and clothes are
taking part in an excellent role within the style and production of ancient fashions with their
autochthonic information for current market demand consumption, mistreatment loom machine
and traditional garment technology. however, their productivity full of shortage of the raw
materials, market problem, poor quality, luck of coaching /education, , technological issues and
money problems. Moreover, the production technology, flexibility and style of the fashions were
not still modernized and not competitive within the international market [34].

The Supply Chain

A supply chain comprises operations including design, manufacturing, marketing, distribution

and customer assistance. Five variables are involved in the supply chain operations. First of all,
the development stage of products is motivated by the demands of fashion and seasons. The
second focuses on the presentation to the world of a selection of clothes. The third is the iterative
process of product development. Fourthly, because of the short life cycle of clothing it stresses
periodic planning and growth. The retail purchasers make the final decision about the finished
items.

The operations of a supply chain have four main management responsibilities inside clothing
companies, namely the planning, organizing, management and monitoring of products, and these
are conducted in each step of the supply chain. The product information is determined by the
designers hired by the shops. This product information are followed by manufacturers in close
contact with the store to build prototypes. The fiber is transformed into a tool with several value-
added operations throughout manufacturing. Upon approval of the clothing prototypes, the
managers in the production companies begin to plan production, which offers directions for
model making. After cutting, the fabric pieces are sewn together with the addition of accessories.

16
 Figure 2. Traditional apparel supply chain. [35]
To work as a designer, you should have an artistic and creative personality. You also have to be
good at drawing and able to express your ideas in sketches. You do not necessarily have to be a
great artist, but you must have some special skills for combining colors, tones and shades.

17
 2.3.Fundamentals of Artificial Intelligence
Artificial intelligence is a broad topic that covers a variety of different activities. Artificial
intelligence can be used in intellectual activities. A.I. is present in our daily lives, and we
sometimes live with it without even recognizing it. It has even proven through time that it is a
developing area with enormous potential to benefit human existence. It is difficult to describe
since it is so broad, with numerous definitions and subfields connected with it. Although it
appears to be a new development, McCarthy, Minsky, Claude Shannon, and Nathaniel Rochester
used this phrase for the first time in 1956 when they conducted a two-month workshop.

AI refers to a large collection of technologies for programming computers to act intelligently.

Artificial intelligence is as much about comprehending knowledge as it is about generating it,
and some definitions of artificial intelligence lean more toward perception and comprehension of
knowledge, while others lean toward a more practical meaning [36].

Table 2. Adapted version of the scheme of Norvig and Russell

18
 2.3.1. Machine Learning
Machine Learning (ML) is a computing field aimed at increasing machine performance via
experience for an activity. In 1959, Arthur Samuel described machine learning as a "discipline of
research that allows computers to learn without being programmed explicitly." But T. M.
Mitchell [8] has been given a more formal meaning, which is “A computer program is said to
learn from experience E with respect to some class of tasks T and performance measure P if its
performance at tasks in T, as measured by P, improves with experience E”.

The subject is at the intersection of a number of areas, including computer science, analytics,
algebra, and engineering, all of which have overlapping components. In the last two decades,
machine-learning approaches were utilized in analytical thinking, predictive modeling,
diagnostics, management, and robotics. ML emerges in a variety of sectors as a significant area
of computing science and a prosperous area of study. In particular, on-line search, safety data,
healthcare, financial modeling, commerce, online recommendations and self-driving vehicles [9].
Depending on the task, there are several methods within Machine Learning.

2.3.1.1. Supervised Learning

In supervised learning, the model is trained given a known input-output pair. The supervisory
signal, which is the intended value, uses each known value as an input vector. Using existing
labels, the technique predicts the labels of the intended output. Classification algorithms involve
supervised learning and may be used to recognize faces, traffic symbols, detect spam in text,
convert audio to text, and a variety of other tasks.

2.3.1.2. Unsupervised Learning

Unsupervised learning is the process of discovering the interrelationships between the elements
of a data collection and then categorizing the data without the use of labels. Two algorithms use
these approaches: clustering anomaly detection and NN. Clustering is a technique for locating
related components or anomalies in a data collection. Anomaly detection using unsupervised
learning is frequently utilized in the security area.

2.3.1.3. Reinforcement Learning

Reinforcement Learning is a methodology that bridges the gap between supervised and
unsupervised machine learning methods. Reinforcement Learning uses both labeled and
unlabeled values as training data. Reinforcement Learning is the middle ground between
19
unsupervised and supervised learning. When paired with a little quantity of labeled data, the
unlabeled data can increase learning accuracy significantly. A variety of scientific assumptions
underpins deep learning techniques. The first is that data with similar labels are labeled together.
The cluster assumption, which asserts that all data in a cluster has the same label, is the second
assumption. The third aspect is that the data is limited to a single dimension rather than being
used across the full input space.

Table 3. Different Machine Learning Algorithms

Classification Clustering Regression

Bayesian Centroid Lasso
Boosting Density Linear
Decision Trees Distribution Ordinary Least Squares
K-Nearest Neighbor Hierarchical Polynomial
Logistic Regression K-Means Ridge
Random Forests Mean shift Splines
Support Vector Machines Stepwise

2.3.1.4. Artificial neural network

A relatively new form of machine learning technique created by associative learning are artificial
neural network and neural nets (NNs). In the logical calculus of the ideas inherent in nervous
activity, the first artificial neuron was developed. These Threshold Logic Units (TLU) or Linear
Threshold Units were designed to represent the operation of a neuron network in the brain
digitally [37].

Artificial neural networks (ANNs) can be viewed as algorithms that extract several
representational layers from input data. An algorithm is made up of several fundamental
computations that, when combined, form a complex algorithm or network. ANNs are arranged
into layers, input layers, output layers, and hidden layers between them. A hidden layer merely
shows that the input or output is not complete. The network is considered a deep neural network
if it has more than one clock layer (DNN) [38] [39].

20
 Figure 3. Artificial Neural Network architecture
Frank Rosenblatt invented the first artificial neuron, which he termed perceptron, in 1957. A
neuron gets one or more inputs "x1; x2…; xn," each of which has a weight "w1; w2...; wn." The
relevance of the input to the output is determined by the weights. Two vectors, x and w,
represent the inputs and weights. The weighted sum is obtained by taking the dot product of the
two vectors. The weighted total is then transferred to an activation function, which decides the
output. This procedure is carried out in each neuron, resulting in a complicated algorithm when
there are many neurons linked to each other [40]. How effectively they generalize to new data is
a key component of all learning models. Real-world modeling will probably not be employed on
data on which it was trained, but rather utilize new data modeling which it never saw before. The
terms used to address model machine Learning performance are overfitting and under fitting.
Under fitting refers to models that do not learn or generalize into new information. It only
implies that the model is not appropriate and something else should be tried on the other side,
overfitting means that a model works well in training data but badly in test or validation data, i.e.
new unknown data. Overfitting could happen if the collected data is insufficient or the
performance of the model is not sufficient. In these particular instances more, data is being
collected and the model capabilities are increased, e.g. number of parameters respectively solves
the problem. [41].

As with other ML approaches, NNs are meant to enhance a task gradually over time. Tasks
linked with NNs typically need a widespread solution other than standard rules-based programs
or statistical approaches, for example computer vision or speech recognition [6].

21
 2.3.2. Deep learning
The word Deep Learning was recently used to characterize networks with "deep" structures,
meaning they contain many more hidden layers than traditional Neural Networks (Figure 4).
Convolutional Neural Networks (CNNs) [42], Long Short-Term Memory (LSTMs) [43],
Recurrent Neural Networks (RNNs) [44], and Variational Autoencoders (VAEs) [45]; are just a
few examples.
Deep learning is a machine learning sub-section that uses deep
(artificial) neural networks that contain more than one hidden layer to
allow machines to understand more abstract data patterns and concepts
in order to do specified tasks. E.g. Recognition of image, self-
captioning, speech synthesis, natural language processing, and so on
etc. [46].
Deep learning is a type of algorithm that employs many layers to Figure 4 How deep
gradually extract higher-level functionalities from the raw input. For learning can be a set of
machines learning and
example, lower layers may detect edges in image processing, but the way ML is a set of
higher layers could detect ideas that are meaningful to human beings, AI [47].
such as figures or letters [47]. In addition, Deep learning can be defined as Computational
models are up of several processing layers that are used to learn data representations at various
degrees of abstraction.

Figure 5. Deep Learning Architecture

22
 2.3.2.1. Convolutional Neural Network
These networks were motivated by studies on the mammalian visual cortex [48], and were first
described by Lecun et al. in their work on optical character recognition [42]. The concept was
inspired by the human brain's hierarchical network of neurons, which is triggered as we detect
items surrounding.
Traditional Convolutional Neural Network (CNN) architecture involves four components :
Convolution: The objective of Convolution is to know visual characteristics from an image such
as edges, lines, etc. such that, after learning the feature on a certain place, it may then be
subsequently identified if it appears at another position.

Pooling: This sample layer (reduces), removes extraneous information and preserves only
needed functions, detects the properties of the preceding convolution layer. This decreases
ultimately the network size and parameters, which are helpful to avoid overfitting the model. It
can be done by either average or maximum pooling operations.

Flattening: Here, pooled feature maps with matrix structures are transformed into column
structures, which will constitute the neural network's image processing input layer.

Fully connected: Neural networks with input, hidden, and output layers are created from
flattened layers. Every neuron in one layer is linked to neurons in other layers in this system.

Figure 6. The architecture of a CNN that performs classification of Animals

2.3.2.2. Recurrent Neural Networks (RNNs)
The simplest form of neural network is a Feed Forward Neural Network (FFNN), which consists
of an input layer, an output layer, and one or more fully linked hidden layers. Recurrent Neural
Networks are Feed Forward Neural Networks in which the hidden layer activation is determined
not only by the preceding layer's input but also by its own activation the last time it fired.
23
This feedback loop enables the network to encode time and sequence dependent information,
which is ideal for dealing with streams of input where the action based on the present state, is
reliant on what happened in the past, such as the meaning of speech, text, and video. One of
RNN's common problems is that the feedback loop causes gradients to disappear and explode
during training and operation. This leads the network too fast, like the problem of the
disappearing gradients that the deep neural network backpropagation faces, to loosen time-
dependent data [46].

Recurrent Neural Networks (RNNs) has dynamic temporal behavior, using loops within hidden
layers to pass information about previous steps to current steps. This allows each loop to have a
memory influenced by a previous history of states.

Figure 7. Recurrent Neural Network with two hidden layers

2.3.2.3. Long Short-Term Memory (LSTMs)
Long short-term memory networks (LSTMs), initially proposed in 1997, and they are one of the
most popular kinds of RNNs [43]. In many different aspects, LSTM are related to RNNs, but in
comparison to more current 'short-term' events, may store long-term input representations. They
were developed to avoid dependence in RNNs that resulted to a disappearing gradient problem.
The models employed the gradients to change their weight over time during backpropagation to
make them unworkable to learn. They feature the same RNN control flow consisting of a chain
of modules with hidden from view internals but with sophisticated gates which may control
information processing. These gates utilize a series of procedures to generate predictions based
on an extensive history for the network to ignore or recall important information. This makes
LSTMs helpful for temporal sequencing applications requiring a longer memory than RNNs.
Include prediction of time series, detection of abnormal time series, language modelling,
sequence-by-sequence learning and creation of music.

24
 Figure 8. LSTM containing four interacting layers example.
2.3.2.4. Variational Autoencoders(VAEs)
An Auto Encoder is a neural network shaped like an hourglass, with the center most hidden layer
being smaller than the input/output layers. The input is essentially stored in a much denser
format in the intermediate layer by training the network to duplicate its input and make the
output as indistinguishable from the input as feasible. A variational auto encoder as specified by
the Auto Encoder but rather than learning a feature representing the data, it learns parameters
reflecting the data distribution (mean and variance). In addition, by introducing KL divergence
as a term in the loss function, it compels the latent space pz (z) to assume the form of the known
distribution q (z) (usually Normally N (0, 1)). This allows encoded data points of similar data to
map each other in points adjacent to the other, producing clusters of similar data (with an
ordinary Auto Encoder that may randomly map points in latent space). This permits the
visualization of relations between high dimension data in a lower dimension space and the
generation of new data by sampling of new data points and feeding them through the generator
portion of the network from the latent distribution [45] [49].
VAEs may be beneficial to several data kinds, including sequential, non-sequential, continuous,
discrete, labelled, or unlabeled, and were utilized for picture synthesis by some of the early
models. While new approaches contributed to complicated generating tasks, VAEs remain an
effective strategy in many issues, including the creation of images.

25
 2.4.Generative Adversarial Network
In machine learning, a discriminative model is one that is generally used for classification. M
Discriminative models take a collection of characteristics X and derive a category why of
whether the picture is of a dog or a cat based on these features. In other words, they attempt to
predict the likelihood of class Y given a collection of features X such as having a moist nose but
not purring, implying that it is most likely a dog. Generative models, on the other side, seek to
learn how to create a realistic representation of a class. For example, a realistic image of a dog.
More specifically, generative models attempt to represent the probabilistic model of X, the
varied characteristics of having a moist nose, provided that a dog's class Y. These models would
create accurate and diverse representations of this class Y with the additional noise.

Generative models come in a variety of forms. VAEs (variational autoencoders) and GANs
(generative adversarial networks). VAEs generally use two models: an encoder and a decoder,
both of which are neural networks. They start by feeding realistic photos into the encoder to
learn. The encoder then has to figure out how to represent the image in latent space. The VAE
now passes this latent representation, or a close approximation of it, via the decoder. The
decoder's objective is to recreate the realistic image that the encoder viewed previously. Indeed,
the variational part injects some noise into this entire model and training process. Rather of
having the encoder encode the picture in one point, the encoder actually encodes the image into
an entire distribution and then samples a point on the distribution in order to feed in the decoder
to generate a realistic image. This adds a bit of noise because this distribution can sample various
points. Generic adversarial networks (GANs) operate in a unique manner. They are made up of
two models once more, but this time there is a generator that creates pictures similar to the
decoder and a discriminator that is actually a discriminative model hidden within. In the name
generative adversarial networks, these models fight against each other, which is why they are
called adversarial. In short, generative models are able to create realistic instances like the
images of adorable pets. An artist who tries to learn how to make photo-realist art is a generative
model. Differentiating classes, such as a cat or dog, are distinguished using discrimination
models. Generic adversarial networks (GANs), on the other hand, learn to pull samples from an
implicitly learnt pmodel. GANs produce photos that are of greater quality than those produced
by other methods.

26
 2.4.1. Common GAN Architecture
Generative Adversarial Network (GAN) [15] is a deep generation model that has been found to
be one of the most prominent in studies of image synthesis such as image editing and generation
of fine grain objects.

The main idea behind GANs is to create a game for two participants. One of them is known as
the generator. The generator generates samples that are supposed to be drawn from the same
distribution as the training data. The discriminator is the opposing player. The discriminator
examines samples to determine whether they are genuine or forgeries. Traditional supervised
learning approaches are used to train the discriminator, which divides inputs into two groups
(real or fake). The generator has been programmed to deceive the discriminator. We may think
of the generator as a counterfeiter attempting to create false money, and the discriminator as a
police officer attempting to allow genuine money while catching counterfeit money. In order to
do better in this game, the fake must learn to manufacture money that cannot be distinguished
from real money, and the generator network must learn how to produce samples from the same
distribution as the training data [50].

Figure 9. Generative Adversarial Networks Architecture

GANs are subject to the maximum probability principle. The fundamental notion of maximum
likelihood is to construct a model that offers an approximation of a distribution of chances
parameterized by parameters θ. The probability that the model assigns to the training data then
refers to the probability: Qm i=1 pmodel x (i); θ; data set with m-samples x (i). The maximum
likelihood principle says the model parameters are selected which maximize the probability of
the training data.

27
To discover a generator distribution pg flowing over data data x, the generator constructs a
mapping function G (z; θg) from a previous noise distribution pz (z) to data space. Moreover, the
discriminator, D(x; θd), returns a single scalar indicating the likelihood that x originated from
training data rather than pg. Both G and D are trained concurrently: we modify parameters for G
to minimize log(1 - D(G(z)) and settings for D to minimize logD(X), as though they were
playing a two-player min-max game with value function V (G; D):

You have to keep the generator scale and the discriminator near to each other to maintain a
successful training process in order to enhance the training over time. And if your generator is
very good, you are done, of course.

2.4.2. Conditional Generative Adversarial Nets

The term "unconditional generation" refers to the process of producing outputs from random
classes. Conditional generation, on the other hand, allows you to request an example from a
certain class and receive it. how to manage the output and receive examples from a specific class,
or how to have those examples take on specific characteristics You can obtain produced
examples from the classes you choose with conditional generation, but with unconditional
generation, you get examples from a random class. As a result, using conditional generation, you
must train your GAN with labeled datasets, with the labels corresponding to the various classes
you need, but unconditional generation does not require labels [51].

Apart from details, CGAN is the same as GAN, and is conditioned by y. “y” could be Any type
of auxiliary data, such as class labels or data from different modalities, may be equating below.
The conditioning can be carried out via supplying the discriminator and the generator as an extra
input layer.

Researches use cGAN to automate texturing of clothing for use in fashion-related applications, to
texture transfer, and virtual try on to replace a person with a new clothes-related apparel [52].

28
 2.4.3. Deep Convolutional Generative Adversarial Networks
DCGANs (Deep Convolutional Generative Adversarial Networks) are a type of GAN with a
unique design that is widely used nowadays. The unique design has resulted in training that is
more consistent and the ability to create pictures with better resolution. The inventors of
DCGAN used modifications to the CNN architecture that had recently been shown [53].

The first approach was to create the network's overall structure comprehensive. Strider
convolutions (discrimination) or transposed convolutions, also known as fractionally strider
convolutions, (generator) with a stride larger than one were used to replace the pooling layers.
The generator was able to learn a spatial up sampling as a result of this. The hidden layers that
were fully linked were eliminated. The generators' input z has been changed, and the
discriminator's final convolution layer has been flattened and fed into a single sigmoid output.
The second method was to employ batch normalization in the layers, which resulted in each
neuron having a zero mean and unit variance. This assisted in learning stabilization by allowing
the gradient to run deeper and preventing the generator from collapsing. The discriminator's
input layer and the generator's output layer, on the other hand, were not batch normalized to
allow the model to learn the proper mean and scale of the data distribution and prevent sample
oscillation and instability [53].

2.4.4. Cycle-Consistent Adversarial Networks

CycleGAN (Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial
Networks) [54] was created to address the problem of the Pix2Pix model, which requires a large
number of tagged picture pairings. And it is handled by employing a mix of adversarial and cycle
consistent losses to translate an image from a particular source domain to another target domain
without having paired instances.

G: XY and F: Y X, and associated adversarial discriminators DY

and DX. DY encourages G to translate X into outputs
indistinguishable from domain Y, and vice versa for DX and F.

Figure 10. Simple Cycle Gan Architecture

29
Researchers applied CycleGAN on several applications where paired training data does not exist.
Like Collection style transfer, object transfiguration, Season transfer, and Photo generation from
paintings, Photo enhancement.

2.4.5. A Style-Based Generator Architecture for GANs

The generator of the original Style-GAN design is made up of two major components: the
mapping network, which is an 8-layer multi-layer perceptron and receives its input directly from
the Gaussian distribution Z., and the synthesis network, which is a convolutional neural network
(see figure 11). The convolutional synthesis network is developed during progressive training,
which implies that a new block of layers is introduced to the generator for each new phase of
training, resulting in output samples with twice sizes [55].

Figure 11. StyleGAN Architecture

The style-based generators can map subnetwork and an intermediate learned latent space W.
Ideally, disentangling the latent space means that the dimensions in the latent vectors correspond
to no more than one feature. The mapping network is a fully linked network that encodes the
original latent vector into an intermediate latent space W, which does not require following the
original probability distribution Z. The style vector is a new vector in the intermediate latent
space that is transformed and incorporated into each block of the generator model after the

30
convolutional layers via an operation called adaptive instance normalization, or AdaIN, which
encodes the w vector from the mapping network into the generated images [56].

2.4.6. Progressive Growing of GANs

Progressive training (ProGAN) was proposed by Karras et al [57], which develops both networks
gradually by starting with a low resolution and gradually adding new layers to both the
discriminator and the generator as training continues. The process is accelerated and stabilized as
a result of this stepwise approach. The network learns the large-scale structure of the pictures at
lower resolutions, and then the learning shifts to finer features at higher resolutions. The
generator and discriminator are mirror images of each other, and they both grow at the same
time.

ProGAN can find low-frequency information and grow more finely scaled high-frequency
information gradually. The generator and discriminator networks develop coordinated, which
implies that the training can train all layers. So that a new layer fades seamlessly, the resolution
gradually increases from 4X4 incrementally to the greater resolution (1024X1024). As the
resolution steadily increases, the generator (G) and discriminator (D) continue to learn at the
same time.

Figure 12. Progressive Growing of Gans Sample

31
 2.5.Overview of deep learning applications for Fashion Industry
AI has a wide variety of applications in the fashion industry through the use of Deep learning,
algorithms, and big data to increase human and market capacities. Such Applications are design
and product creation [8] that help new product creation by analyzing design components such as
color, cloth, shapes, and cut. As a designer [9], AI creates innovative fashion designs along with
sewing templates by using a powerful algorithm that analyzes past designs and future trends. AI
in textiles can be used in clothing production by automating cloth quality management, pattern
checks, color matching, and defect identification. From the point of view of purchase and
merchandising [10], AI used in predictive analysis by automatic product labeling and pattern
predicting, personal styling platforms, visual merchandising platforms, and performance
measurement. Online product and visual resemblance recommendations complete the look
recommendations, and customized recommendations per country, per user section, and per
individual customer are all examples of AI in e-commerce [11]. AI can be used in automated
product tagging, natural language search, visual search and chat-bot support for customer care
and clienteling, branding strategies, and marketing strategy [10] [12] [13] [14].

Table 4. Artificial Intelligence Application in Fashion Industry

32
Technology has always been an important factor for automating fashion, without any doubt.
However, the digital transformation is the spread of digital technology and the wealth of data
created by these technologies. The fashion sector began a more thought-provoking and quick
change, which varies how customers buy and interrelate items and brands. Fashion & Textiles
uses these technologies to better meet their objectives in the supply chain or employ real-time
data and analytics in order to better anticipate demand and increase prices. In recent decades,
research into artificial intelligence (AI) and its implications for the textile and fashion industry
has become a much-discussed issue and gained more attention from academics and industrialists
as a consequence of the un-limiting growth of computer technology [58]. The applicable usage
cases have been widely and severely changed by derived AI and associated technologies. The
resultant efficiency and efficiency of these applications has drawn additional attention and
motivation in other fields, such as virtual fashion advisors and stylists; market analysis and
predictive analysis; automated client care and chatbots; and so on.

2.5.1. Deep Learning Fashion Image Generation

Deep generative technologies have led to the successful production of images/videos, some to
style transmission, and some to a composition of specified conditions [52]. Deep generative
models such as the Pixel Recurrent Neural Network (RNN), autoencoder (AE), variational AE
(VAE), and other forms of Generative Adversarial Networks (GAN) have sparked interest in
creating or reconstructing samples based on existing datasets.

AI is predicting the performance of the sewed seam, garment design, other sewing processes and
quality control more often. AI may be used for the intelligent production of clothes in order to
forecast the characteristics of clothing following a certain technique. It may therefore be utilized
for appropriate clothing design through manufacture and monitoring of the processes of
production. Based on seam puckering, seam floating, and seam efficiency, (Hui & Ng, 2009)
used ANN to study the performance of textiles during garment production. Despite using a little
number of training data, the ANN model proved extremely accurate in detecting fabric defects
[59]. In order to select the best appropriate manufacturing facility for a certain client order, an
AI-based decision support system was utilized in decision making. It can also help with
production planning and control.

33
Zhu et al. [60]trained a two-stage GAN to create new equipment’s based on textual description in
a given picture, thereby making structural coherence even more consistent with appearance and
position. However, its dataset and segmentation maps (128 x 128) constrain resolution since
further information is difficult to be obtained. Gu ̈ nel et al [61]. utilized the conditioning
mechanism (FiLM) to combine visual depictions and textual representations but therefore caused
some deterioration without utilizing any spatial restrictions to meet the same objective In
general, the look and pose, form and texture of the clothes are two categories of fashion
components. Some of these aspects must be disentangled or not altered at the same time,
depending on certain application circumstances. Li, Y., Yu, X., Han, X., et al.\ 2020 [62]divided
the job of synthesizing garment images into texture and shading, proposing an interactive fashion
image synthesis sketch method. The technology enables users to create contours, provide various
texture information and increase shading by designing shading borders. To transform this spare
information into realistic photographs, deep learning algorithms are used. A new bi-colored edge
representation is suggested to balance output quality and interaction effort. The study by
FashionQ [63] investigates how AI may assist creativity and identifies key factors to consider
when building AI-based creativity support tools (CST) in the context of ideation in the fashion
design sector. Creativity is a subjective notion that varies according to persons and surroundings.
In this project, they worked with fashion designers to learn about their existing design processes,
aims, and difficulties. They built and developed an AI-based CST with fashion designers through
an iterative approach that externalizes three cognitive functions extending, constraining, and
blending in overcoming design fixation throughout the fashion design ideation process.

One of the main usages from Deep Learning Fashion Image Generation is style transfer. Style
transfer converts an input picture to a matching output image for example transfers a real-world
image to a cartoon image, or transfers the clothes tested on a human image, from one style to
another. The image processing transfer style includes a broad variety of applications, such as
face make-up and virtual reality [64].

AI may also be used to detect and evaluate textile fibers according to color and other
characteristics, such as fine, longitude, uniformity ratio, toughness, and spinning effects [65]. and
AI used to adjust the aspects of garment look, comfort, and durability while meeting fabric
requirements.

34
 2.5.2. Deep Learning for Fashion Recommendations
Not everyone is a natural fashion stylist, the fashion recommendation, given its obvious
applicability for Online Shopping in fashion items, has received ever more attention to support
this requirement. [64].Fashion recommendation works based on Fashion compatibility that
determines how effectively different types of items can work together to create attractive attire.
Fashion suggestion considers both aesthetic resemblance and visual compatibility across various
clothing kinds.

Convolutional Neural Networks (CNNs) have seen a renaissance as generally effective image

models, notably for personalization [66]. Such visual measures were shown to be useful in
interface expectations, for example proposing electing things (for instance two comparing shirts)
or reciprocating things (for instance, a shirt and a pair of coordination jeans). Chen et al [67].
Used a CNN model to create an image-based product recommendation system that included
product category identification and similarity calculations. CNN, a deep learning model, has a lot
of potential for developing knowledge-based apparel recommendation systems because it can not
only extract low-level features like SIFT and HOG, but it can also train a feature detector to
recognize highly sensitive clothes features and adapt specific classification tasks. As a result,
CNNs can reach greater levels of picture feature extraction accuracy.

When it comes to outfit matching, each outfit usually consists of several complimentary pieces
such as shirts, trousers, shoes, and accessories. The matching fashion elements are crucial to
putting together a beautiful ensemble. However, for a variety of reasons, producing harmonious
fashion matching is difficult. First and foremost, fashion is a delicate and subjective notion.
Second, there are several qualities that may be used to describe fashion. This topic has sparked a
lot of attention in recent years, resulting in a lengthy number of algorithms and methods.

Al-Halah et al., 2017 [68] are proposing a model that unsupervised manner discovers fine visual
styles from large-scale fashion data. Their approach finds distinct style signatures and describes
them semantically based on the main visual features for each one. In addition, their model also
predicts the future acceptance of styles and shows their life cycle and status based on consumer
consumption behavior (e.g. in- or out of fashion). They show that vision, which outperforms
textual representations, is crucial to trustworthy predicts. (Yu et al., 2018) [69] Proposes
incorporating aesthetic information into clothing recommender systems, which is extremely
35
relevant to user choice. To do so, they looked into the use of aesthetic characteristics for
customized recommendation using implicit feedback datasets, then presented the aesthetic
features extracted by a pre-trained neural network, which is a brain-inspired deep structure
trained for the aesthetic evaluation job. They offer a novel tensor factorization approach to
include aesthetic characteristics in a customized manner, given that aesthetic preferences vary
considerably from user to user and over time. Difficult real-word datasets experiments indicate
that the approach offered substantially surpasses state of the art models and successfully
recommends goods, which suit consumer styles.

AI Personal Stylist: Having a personal stylist would be the pinnacle of customization, but it
would be beyond of reach for the ordinary individual. This is feasible because to AI, which has
enabled the creation of personal virtual AI stylists. This product represents the confluence of all
of the previous technologies: natural language processing, natural language comprehension,
computer vision, neural networks, and different forms of machine learning. The Virtual Style
Assistant is a step ahead, as it highlights the use of images, the opportunity to take photos, the
ability to recognize images and search visually as well as the engines for recommendations and
accessibility to fashion goods [70]. Fashion luxury brands in connection with their actual stores
use intelligent mirror technology. Users may utilize the smart mirrors to not only view how they
appear in different colors of clothes, but also to alter clothing types and compare ensembles side
by side. This makes buying much more pleasurable and convenient.

The Beauty e-Experts system was later created by Liu et al. [71] To automatically propose the
most appropriate face haircut and cosmetics, as well as synthesize the visual effects. It was
suggested that the removed facial and clothing properties should also be used to simultaneously
learn multiple super-graphs with tree structure, so as to model in collaboration the relationships
underlying the attributes of beauty (e.g., hair length and eyelashes color), the beauty of mid-level
(e.g. shape of the eye and mouth width) and the image of low level. In addition, a module to
synthesize the qualities of beauty indicated by the multi-tab structured super-graphs model was
also suggested.

The retailers additionally manage merchandising customization using AI Recommendation

Engines. They provide customized products based on client data and are typically proposed with

36
phrases like "you may like too." Engines for recommendations let consumers sort huge volumes
of data they do not have to provide.

2.5.3. AI for Fashion e-commerce

With the availability of e-commence platforms and the convenience of internet shopping,
customers take on a wide range of different alternatives. Purchasing is one of the most significant
activities in our society, and it is vital for the economy, sociology, and psychology to research
consumers’ shopping habits. One of the most popular areas in e-commerce is undoubtedly
clothes. The study of popular and appealing clothing aspects has been needed, and many
developing applications, including clothing recommendations and advertising, might further
enhance [72]. Because of the growth of online purchasing services, the apparel sector is one of
today's fastest expanding businesses. Consumers' frequency in the use of broadbands and mobile
devices, changes in internet content, future experience in online buying and the continual
upgrading of the purchasing procedure and comfort are among the contributing causes for such
huge development. Online shopping is prominent in apparel among the most popular categories
of e-commerce. The manual observation would be the conventional technique of detecting the
trend of clothes vending and positive design features. However, it takes a lot of time and will
change with the time to time. One major problem of Fashion is that the production quantity per
design is less compared to the customer’s demand. Today’s ‘fashion is accompanied by a process
of continuous innovation, in which new designs are developed at sometimes big cost only to be
replaced by other designs. [73].

There are numerous areas in retailing where AI application is expanding, customer relationship
management, including sales forecasting, fashion retail forecasting, assessing customer
happiness, style advice to customers, demand forecasting, and fashion coordination.

Without AI applications, it would be hard to handle the rising scale and complexity of
personalization in digital way. Chatbots or AI smart assistants are the most popular solutions for
personalized online purchasing. Virtual machines engage with consumers through chat, reply to
customer care requests, and assist users to traverse the web and shop sectors, advice apparel and
accessories that suit a particular client as if they were people who work 24/7 a day. Artificially
intelligent chatbots, on the other hand, can comprehend human language and respond to non-

37
predefined inquiries. Customer satisfaction with live chat is greater (73 percent) than with phone
or email, according to studies [74].

2.5.4. AI for Fashion Virtual tray-on

Virtual fitting systems bridge one of the greatest gaps between e-commerce and retail. The goal
is an image new fashion picture synthesizing by smoothly merging diverse elements. These
technologies may be used in potential applications from item retrieval to skilled fashion design
helpers. The main objective for VR technology is to generate a sense of reality that makes it
impossible for users to discriminate between actual and virtual environments. Remote control
technology based on robots have drawn increasing attention in society by continuously
improving human life and industrial production requirements. And in many areas, this
technology has played a very important role in the development of relevant research and
technical growth.

In the field of display design, it has a very major practical relevance based on the use of artificial
intelligence and VR technology. Clothing is peculiar since it is not utilized alone. The matching
of apparel accessories is therefore a significant component in the process of clothes sales. Good
clothes display may be a type of silent education, not simply for buyers. Customers do not have
to display a bad attitude so that they may learn how to match apparel. The exhibition shows the
beauty of its clients directly and transmits the artist's ideas through this neutral medium. There is
no anybody in the crowd any more. If you do not have self-emotions, like envy, you will more
likely embrace a concept. When the buyer accepts the notion, he agrees, of course, that the
product is actually lovely and simpler to purchase [75]. The aim of display is to present to our
customers all the material and emotional characteristics of goods and brands using creative
methods and design techniques as well as other visualized languages, to create a product image
in the customer's minds, to generate interest, preference and confidence in products and brands,
thereby exciting the needs and wants and desires of our customers.

In luring consumers into the retail shop or visiting a website, visual merchandising has a vital
role. In the store and on the Internet, visual merchandise is increasingly utilized to improve
consumers' buying experience and to extend the accessibility of the brands. Visual
merchandising displays are very influential for customer shopping. The manner that goods are
marketed and presented in the shop may have a significant effect on the behavior of consumers.

38
According to the findings of several studies, visual merchandising of intimate clothing has a
modest influence on female customer purchase behavior, since the majority of female consumers
dislike the usage of mannequins for lingerie display. e-commerce provides a judgement-free
environment in which to acquire the goods, as well as a wider range of lingerie products in terms
of design and color [76].

The acquisition is visual. Looking at is what is purchased. Brands have acknowledged this worth
and promote themselves to draw the interest of the customers. Displays must only appeal in
visual terms, to engage and motivate the client. Consumer behavior is also visual and tactile and
senses are used for purchase. The desire to see, touch and when everything is pleasant to the
impacts of a purchase have maintained marketers and retail experts investing in visual
professionals, engaging more customers. Visual merchandising facilitates the location and self-
selection of a product for a consumer. Some displays show the customer how to wear a product
in the event of clothing and how to utilize a product. It is also the technique to move the items of
slow sales – to make them enviable for the consumer and they choose it off the shelf or from the
web shop [77].

Figure 13. Virtual tray on in store.

AI will certainly drive the fashion industry to a more sustainable environment, expand business
models for retailers, and make them more affordable. AI is also expected to radically disrupt the
fashion industry, because of not only new business models and production processes, but also
because the individuals who work in the fashion industry and their vocations will undoubtedly
have an influence. Artificial Intelligence (AI) is sometimes considered as less serious. Important

39
fashion figures. AI has the potential to progress in many sectors of the value chain, and others
will follow. We can use AI for selection with a lot more data science. This might extend all the
way to the back-end infrastructure and production. It does not entail the greatest product range; it
concerns providing customers with the most apparent diversity.

AI can provide customers a customized buying experience: With the new marketing approach,
all companies ensure that they have a better understanding of consumer preferences which will
enable Fashion brands to minimize waste and the number of unsold items.

The construction of 3D models of fashion products has several applications to clothing business
in the fields of virtual reality, mixed reality and Computer-Aided Design (CAD) for clothing-
design.

A lot of study has been done to create virtual model 3D clothing. Some employ unique multi-
camera configurations to record 4D changing clothing form [78]. Other techniques are used as
input for 2D sewing patterns [79] or 2D sketches, which are easy to produce 3D input models.
Another set of 3D template molding technology designs textile clothing that fits finest 3D human
virtual models [80]. Image-based virtual try-on has also been considered as an alternative to 3D
modeling. Neverova et al. have been using the two-stream network, combining a predicted image
driven by data and a tangled surface picture, while at the same time learning from one end to the
other to produce a new personal position [81].

When we think of AI for Virtual Tray on, we also need to investigate Semantic segmentation.
Every pixel is allocated with semantic annotations Semantic segmentation. CNNs were used for
this assignment successfully. For semantic segmentation, Long et al. introduced Fully
Convolutional Networks (FCNs), which resulted in major advancements on handmade
techniques [82]. Built on FCNs, the architectures of the encoder-decoder were a fantastic result.
Typically, such an architecture includes a function map encoder and decoder, which maps the
information encoded to the resolution of the entry. In order to gain a multi-scale level of
information spatial pyramid pooling (SPP) may also be used in many scales [83].

Given a source picture and a 2D-target position, image-based techniques create an image at the
target position with the source appearance. It is useful to change the pixels into an original
picture to match the target posture within the network in order to cope with pixel misalignments.

40
Similar principles of developing GANs may be used to produce high-resolution anime characters
[84]. Synthesis of try-on and conditional clothes. There are an increasing number of recent
papers on making the picture creation conditioned on a specific piece of clothing or a
description. Additional works have shown models that can exchange appearances among two
separate subjects [57] [85] [60].

2.6.Generative Adversarial Networks for Fashion Industry

The image generation task becomes simple after the introduction of the generative adversarial
network [15]. There are two types of image generation technology: generative parametric and
adversarial approaches [16]. Several studies have shown that adversarial approaches, particularly
the generative adversarial network, have clear advantages in image generation (GAN). GAN is
type of machine learning in which two adversarial models are trained at the same time: a
generator ("the designer") that learns to generate images that appear real, and a discriminator
("the design critic") that learns to discern real images from fakes. During preparation, the
generator improves its ability to create realistic images, while the discriminator improves its
ability to detect fakes. This is used dynamically to produce computer-generated images and
motion that looks plausible (and possibly aesthetically pleasing) to the consumer [17].

The introduction of Generative Adversarial Networks (GANs) in 2014 by Goodfellow et al. [15]
is one of the most cited articles in the subject of Generative neural networks. Many other
contributions to the subject have been made since then, which have had a significant impact on
this research. Deep Convolutional Generative Adversarial Networks is the first of these
(DCGAN), Radford et al. [53] presented it, and it turned out to be a good candidate for
unsupervised learning. Both the generator and the discriminator use a convolutional technique.
Their research revealed that the deep convolutional neural networks were used to give learning
that is more robust. Recently, Style based GAN, CGan( generative adversarial networks) and
CycleGAN image-to-image have shown useful when producing photo realistic pictures, the use
of image synthesis models normally involves appreciated characteristics and may fail in the
individual modification of the target attribute/object [52].

GAN’s for Fashion Image Generation and Recommendation

41
Editing attributes has become an important and emerging topic of computer vision (tailor gan).
Given an A image of the reference garment and another B image with the target attribute (neck /
sleeve), generate a photorealistic image that combines the reference texture A and the new
reference attribute B. Generative adversity networks have pretty much achieved a remarkable
success in synthesizing realistic pictures for numerous applications. Therefore, we can use this
technique first by utilizing an encoder Network to convert the query image into a latent Vector
representation. Then this representation coupled with another vector that characterizes user style
preference is taken into the input as for the generator Network that generates the target item.
Therefore, the basic technique is as follows: the goal of customized fashion design is to create a
fashion item for a specific person based on an input query item. Therefore, there are two
fundamental requirements for this design that they have: the first one is the realness requirement,
which literally means that the design item should look authentic. The second condition is
compatibility, which basically means that the design item should be compatible with the query
item [86].

Zhao et al. proposed a deep Nested Adversarial Networks for Fashion detection which includes
three Generative Adversarial Network (GAN)-like sub-nets that can cooperatively learned in an
end-to-end training approach for semantic saliency prediction, instance-agnostic parsing, and
instance-aware clustering respectively [87]. Inspired by CycleGAN architecture, Chang et al.
[88] proposed the idea of asymmetric style transfer and a framework for training both the
makeup transfer and removal networks concurrently, each one reinforcing the other. More recent
work by Chen et al introduced BeautyGlow that decomposed the latent vectors of face images
derived from the Glow model into makeup and non-makeup latent vectors. In [66] they
introduced a framework for style proposal that's delicate not just for proposing current items to a
customer, yet which is likewise fit for developing fresh, plausible design images that harmonize
customer preferences. They utilize Generative Adversarial Networks to become familiar with the
diffusion of style photographs and build novel fashion products that enlarge clients' preferences.
This structure can be utilized in an assortment of proposal scenarios, to examine the space of
feasible fashion products, to adjust existing, or to make items made-to-order to a person.
FashionGAN [89] provides a garment design methodology for image-to-image translation from
drawing to image. Inputting the appropriate fabric image and fashion sketch permits the
development of an image depending on the color of the drawing material.

42
Recommendation System often is considered as the basis that compiles data about individual
desires and taste of a consumer and offer ideas on product sales as per their interests. The so-
called recommendation system takes the information from the user through the following two
techniques: either expressly or impliedly. Indirect way of gathering information entails careful
observation or monitoring of customer's bought items whereas implied acquisition of consumer
preferences involves observation on past ratings or histories.

Building effective recommendation systems frameworks for sectors like fashion is difficult
because of the considerable quantity of subjectivity furthermore, the semantic multifarious
character of the highlights that are in question. Due to the particular aspects of the fashion
domain, most traditional recommendation systems are not suited for use in this domain. For
example, while purchasing clothing, individuals do not naively follow trends or purchase the
same stylish item repeatedly. Another factor is the concept of complimentary relationships when
it comes to proposing a tailored fashion ensemble. Humans have a natural tendency to form
associations between items based on their visual appearance. GAN-based models have recently
shown promising performance in the area of outfit suggestion, competing with state-of-the-art
fashion recommendation models. Finally, another interesting GAN application is using the
generating capacity of GANs to create realistic-looking fashion clothing. This feature may attract
the interest aesthetic or appeal of customers and designers, encouraging them to explore the
range of possible fashion designs.

2.7.Related Works
Image processing is a specialized area of research that deliberately focuses on the perception of
digital images by machines and computers. Since advancing classical computer vision through
sophisticated machine learning methods such as deep learning. It attempts systematically to
simplify a human visual systems mission. Today, it is necessary to apply Deep learning to solve
issues in art and fashion design. This has the ability to radically alter how people communicate
with fashion. Because of fashion subjectivity and personification, e-commerce sites have
excelled at making clothing shopping simpler by search items. Many fashion companies have
researched several similar questions for the huge potential of industrial deployment as hot
research topics [90]. The top thematic research fields in design creation and manufacturing are
now fabric segmentation, recognition, retrieval, recommendation [61], [91], [92], and image

43
production [93] [94]. Likewise, generating high-quality pictures from text descriptions is an
active research trend in the creation of images and has created great fashion interest [95].

GAN [15] has continued to build interest in the deep-learning community since its launch in
2014; it has been used in different fields such as computer vision, natural language processing,
image and video synthesis [96] and semantic segmentation [12]. In particular, in the area of the
computer vision field, GAN has shown considerable recent progress in many tasks such as the
generation of images [13], image to image translation [97], and the super-resolution of images
[67]. Study into image production of clothing and designs much analysis has been conducted
from the point of view of the user, such as the pictures of apparel segregation [98], trend
forecasts, virtual tray-on [99]. In the other hand, the manufacturing process from cutting to
producing the real garments out of the images is not investigated well [100].

From an image-to-image translation perspective, which means the transition of an image from its
original form to any fake forms while preserving the original structure and semantics [12], Many
experiments have been performed to display image translation through several contexts, such as
conditional Gan, CycleGAN, Progressive Growing Gan and Style-based Generative Adversarial
Network. The conditional generative adversarial nets (CGan) [101] is an extension of a primary
GAN through limited conditions, for instance, class information, text descriptions, etc.
conditional Networks use a conditioning vector as the input array of the Network. Upon training,
the user can manipulate this variable to monitor the result generated. CycleGAN [21] is a new
improved method and design model of Pix2Pix [102] model which constructed through a loss
function and training mapping. CycleGAN improved this by translating an image from one
source domain to another target domain without using a combination of adversarial and cycle
consistent losses. Recent research on Progressive Growing Gan indicates that it can aid in
training stability and produce the best results when producing reliable images and again it helps
to generate high-quality images in a progressively growing discriminator and generator [57].
Style-based Generative Adversarial Network (Style Gan) [20] is the result of a synthesis of style
transfer literature that can start from constant input and monitor image synthesis to generate
high-resolution images of higher quality. In addition, Deep Convolutional Generative
Adversarial Networks (DCGAN) architecture utilizes Deep Convolutional Neural Networks to
produce images from a particular noisy data distribution using the generator/discriminator

44
paradigm. In addition, demonstrate persuasive proof that knows the hierarchy of representations
in both the generator and the discriminator from object pieces to scenes [53].

Text-to-image synthesis in fashion is an intriguing application of fashion synthesis [67] to date;

only a little work has been carried out on a text-to-image synthesis for GAN fashion design. Zhu
et al. [60] Trained a two-stage GAN to create new outfits onto a given image based on textual
explanation, resulting in improved structural coherence, including coherent appearance and
posture. Even so, their data set and segmentation maps limit the resolution (128*128), since extra
detail is expensive. In addition, zhengrong yi [103] conducted research on Text-to-Image
Synthesis for Apparel Design. By using two GAN dependent algorithms, AttnGAN and
StackGAN, they propose that AttnGAN shows greater strength in synthesizing images of high
quality for the most common categories, which contain much longer descriptions. The neural
network generation model for producing high-quality images of fashion objects based on text.

Ashenafi Workie [4] has tested and analyzed with more GAN models for an immersive fashion-
modeling framework with visual awareness. The architecture is designed to combine current
ProGAN architecture with controlling inputs such as color, shape and texture. The ProGAN
generator takes the inputs and preserves images generation with a constraint, while the
discriminator proceeds to distinguish images in real and false terms. The test results of the
proposed solution show that fashion design can be achieved by manipulating significant
characteristics during the generation using a profound generative model algorithm. However, the
proposed approach is only applicable to color. The fact, on the other hand, requires numerous
and complex texture or pattern, scale, printing form, and text. In addition, the model must create
a high-quality image that accepts several conditional inputs.

Garment Design with Generative Adversarial Networks implements Design-AttGAN, a deep

learning algorithm that can automatically edit garment images based on certain user-defined
attributes. To prevent the implicit discrepancy between the reconstruction loss and the attribute
classification loss, the initial formulation of the AttGAN [104] is updated. From this research, we
may observe that generative adversarial networks are vulnerable to various domains,
necessitating cautious revision and hand engineering of the algorithms depending on the dataset
and goal mission.

45
Recent researches on Fashion image synthesis using generative adversarial networks shows some
advancement. Even though GAN has shown considerable and effective outcomes in the past, it is
still a difficult issue for some reason. The gap seen in previous researches are the quality of the
generated images is poor and unable to provide the control of fine-grained styles and not
considering some conditionals to be included in generated images such us, visual artifacts
patterns, text, a huge set of colors and diversity in their datasets, and not included the
perspectives of professional fashion designers.

46
 Table 5. Related works in detail

Author Title Methodology Objective Gap

Y. R. Cui, Q. FashionGAN: To An image synthesis pipeline that can Automatically generated The model does not generate
Liu, C. Y. Gao Display your fashion generate garment images based on a fashion image and new Fashion design image but
and Z. Su (2018) design using fashion sketch and a specified fabric establishes a bijection simply implement virtual
Conditional Generative image using CGAN to show image. relationship between garment display. In addition,
Adversarial Nets. generated images. workable only for single-color
and regular fabric pattern.
Shizhan Zhu et Be Your Own Prada: Given a person's input image and a To present a novel and Not considering using Design
el. Fashion Synthesis with text, that describes a desired outfit, effective approach for sketches. Moreover, training set
Structural Coherence And then the model output the generating new clothing on contains images mostly with
person with a new dress, while a wearer through generative a plain background.
maintaining the wearer's and pose adversarial learning.
unaltered.
Yuan, Mohsen Garment Design with Attribute GAN (AttGAN) which is a To automatically edit Lower resolution and quality of
Moghaddam Generative Adversarial generative model proven successful garment images based on the generated image. In addition,
(2020) Networks for attribute editing of human faces certain user-defined It does not consider many
is utilized for automated editing of attributes. conditionals except shape and
the visual attributes of garments. color.
Ashenafi Workie GAN-based visual- By using ProGan and as input it To design and develop an Not Considering the main
(2020) aware interactive takes the masks of the images and interactive fashion design Conditionals like Pattern, Shape,
fashion design adding the conditionals into the framework texture and not huge set of
masks. In addition, this research try colors. In addition, The
47
 framework to investigate other GAN’s generated image has low quality
architectures applications to generate and the model outputs
fashion images. inconsistent results.
Lele Cheny, TailorGAN: Making By using two methods, those are To investigate the problem Not Considering the Design
Justin Tiany, User-Defined Fashion reconstruction-learning step, which of doing image editing to a sketching process from scratch
and Guo Li Designs the model learns texture and location fashion item with user rather than it manipulates from
(2020) information and adversarial learning defined attribute. In the garment image. In addition,
step which has the capability to addition, To generate photo Does not contain diversity in
achieve single-attribute realistic images with desired their datasets.
manipulation. attribute.

48
Chapter 3. Methodology
3.1.Overview of Methodology
Research is a systematic method of gathering, analyzing, and interpreting data in order to
understand a topic. Furthermore, research is described as "a thorough study of a topic,
particularly in order to find (new) facts or achieve a (new) knowledge". Some features of a study
include that it begins with a topic or issue, involves a strategy for progressing, needs gathering
and analysis of data and is, by its nature, cyclic. Research instruments are used to gather, modify,
or analyze data and a study methodology is used to perform a research. It is essential that the
research technique be appropriate for the selected issue.

Three popular research technique approaches are the quantitative, qualitative, and hybrid
methodologies. They are tailored to the kind of data required to answer to the research question
or address the research issue. The quantitative approach is based on the measurement of quantity
or amount. It handles numerical, non-descriptive, data where statistics or mathematics are
typically used. The qualitative approach is mainly utilized for exploratory research. It is a holistic
method that includes exploration and is used to comprehend reasons, views, or motives. It offers
insights into an issue and manages non-numerical, descriptive, data. The mixed approach is a
"mix" of both quantitative and qualitative. For instance, a qualitative approach may be utilized to
comprehend the meaning of quantitative data.

The following section describes the methodology used in each phase of the project and how it
relates to the project goal to formulate a conceptual framework for generative Fashion design
based on deep learning.

This chapter starts with a description of how the required knowledge and skill are obtained and
concepts are developed. Training. Next, it states the description of how the data is collected,
processed and prepared before training. Then, the next chapter stated the software and hardware
tools used in the successful accomplishment of research. It ends by explained the evaluation
methods and metrics to assess the result of the study.

49
 3.2.Concept Development
The purpose with the concept study is to develop a knowledge of the subject at hand and what it
can be utilized for today and in a near future. This involves recognizing common methodologies
and methods, their benefits, drawbacks and applications as well as problem - solving situation,
challenges and solutions to solve them. This material is gathered through looking for and
analyzing research papers, articles and literature on the topic and by following suggestions and
citations from prominent personalities in the area.

The first step to take after identifying the field of interest is to find best online class to be
familiar with the subject. In addition, the most detail explanation and good lectures is found on
Coursera by DeepLearning.AI. It offers an interesting introduction to image creation-using
GANs, tracing a route from fundamental ideas to advanced methods via an easy-to-understand
manner. It also addresses societal consequences, including bias in ML and the methods to
identify it, privacy protection, and more. Build a broad knowledge foundation and get hands-on
experience in GANs. Train your own model using PyTorch, utilize it to generate pictures, and
assess a range of sophisticated GANs.

Publications is mainly looked for via Google’s search engine Google Scholar and through the
white paper repository ArXiv. Relevant topics are selected from previous searches from these
sites as well as latest trendy topics within the area of deep learning published on. Cutting edge
research is also found by monitoring news and publications from groups such as OpenAI and
DeepMind.

Figure 14. Concept Development Process

50
 3.3. Dataset Preparation and Preprocessing
The suggested architecture objectively should be by utilizing a fresh custom dataset. This is
intended to be done by gathering Ethiopian traditional dress fashion products images and fashion
design sketch images. This custom Dataset collected from the following sources:

 Ethiopian Fashion Design School

This method of collection dataset is a collection of 500+ images with different sizes. The images
are of design sketches from different fashion design school and universities that are the student’s
projects and stored files by school which is used for teaching. The collection contains a lot of
noise in form of cutting, pattern, unusual background.

 Web Scrapping

Gathering and preparing training data is a tedious task: relevant data needs to be gathered,
properly labeled and cropped to suitable size. the data needs to be either created or gathered,
cleaned and preprocessed. In the case where gathering of a vast amount of data is required, it
takes a long time to gather manually. Then an automatic gathering system utilized to simplify the
process. To automate the process of collecting training data this study present a system, based
around a web scraper and an object identification model.

Figure 15. Web Scrapping Process

51
The web scraper generates huge collections of images from chosen sources based on suitable
keywords. Various search keywords on multiple languages could be used to guarantee a variety
in the results and to vary the results to include different traditional clothing. The purpose is to
gather as much relevant existing data as possible from relevant public sources. After collecting
the data is sent on to the manual filtering process which tries to identify whether the image is a
false positive or not. The method is also capable of identifying in which area(s) the object(s) of
interest is/are positioned, making it feasible to trim these pictures to the region of interest.

 Custom Capturing
Taking a picture of the wearable fashion product from different stores around Addis Ababa.
The statistics referenced in this experiment are displayed in the following Table:
Table 6: Dataset Description

No Source Count

1 Custom Capturing 2500

2 Web Scrapping 12000

3 Design School 500

Pre-processing Techniques

After gathering the required data, the following stage is conducting pre-processing on the
obtained face pictures to eliminate some distortion, resizing, and filling the content to make them
with same dimension using various methods.

Clean background is important to display the core portion of desired product in the picture. Due
to this, all training pictures had the clean background. For removing background of images this
research utilizes different API’s that are suitable for this purpose. E.g. clapcut, removebg ... etc.

52
 3.4.Loss Functions in GANs
3.4.1. Cross-entropy loss
The term "log loss" refers to the Cross-entropy loss. It's used to classify things. It assesses the
effectiveness of a model whose result is a probability vector that sums up to one the expected
possibilities of all classes to identify. This may be accomplished, for example, by using a soft-
max function to activate the final layer. This loss is used to classify an image of class A. If the
model assigns a high probability to class B, the loss is not penalized. Instead, it penalizes class a
more for the model's low likelihood. If the likelihood provided to the right class is smaller, the
loss penalizes more in other desires.

Cross−entroppy=−∑ p ( x )∗log 10 q ( x )
x

Where p(x) = possibility of class x in target. q(X) = possibility of class x in prediction.

As the loss just pays attention to the expected possibility. The other terms do not need to be
computed since they will be 0. This could be accomplished if our goal is a one-hot vector. by
considering just, the loss for the hot class. This is known as categorical cross-entropy loss.

Categorical−Cross−entroppy=−log 10 q ( x )

3.4.2. Binary Crass-Entropy

There are two types of binary crass-entropy: zero and one. Using a sigmoid activation function,
the output is a predicted value that is shifted between zero and one. To maintain the loss, we one-
hot encode the two classes (1.0) and (0, 1) and convert the predicted value to a probability
vector. Allocating the likelihood of one minus the forecast to the other class if the goal is 0 or
one, the first and second terms of the formula cancel out.

binary crass−entropy =−¿))

3.4.3. Adversarial Loss

We need to go back to the BCE to understand how the adversarial loss works. The discriminator
works as a binary classifier, attempting to distinguish between genuine and false pictures. When
a real picture enters the discriminator (p(x) = 1). The second phrase canceled out by the first one.

53
Similarly, when a generator-generated false picture (p(x) = 0) is obtained. The first term is
canceled.

binary cross−entropy (D ( x ) , 1)=log 10 G ( x )

binary cross−entropy (G ( x ) , 0)=log 10 ( 1−D ( G ( Z ) ) )

Thus. The objective of the discriminator is to maximize these two functions.

L ( D )=max [ log 10 G ( x ) +log 10 ( 1−D ( G ( Z ) ) ) ]

The generator is competing against the discriminator. Trying to minimize the above equation.

L ( G )=min [ log 10 G ( x )+ log 10 ( 1−D ( G ( Z ) ) ) ]

Putting together the objective functions of both generator and discriminator, the final GAN

Objective is:

L=minG max D [ log 10 G ( x ) +log 10 ( 1−D ( G ( Z ) ) ) ]

The mini-max loss game maximized the generator loss in the early stages of training in the initial
Setting. This necessitated re-framing the concept such that both the generator and the
discriminator maximized a loss function. To do this, the generator method is modified, and
instead of minimizing the discriminator goal, it increases the likelihood that a discriminator
would mistake a false picture for the real one.

The discriminator is a binary classifier that attempts to maximize the binary cross entropy loss.
The generator learns to generate samples with a low estimated likelihood of being false. This
implies lowering the chances of the discriminator incorrectly classifying synthetic pictures as
synthetic.

3.4.4. Losses from Additional Generators

In principle, standalone adversarial training may help the generator understand the data
distribution of the data it has been given. Many GANs settings aspire to understand a particular
transformation rather than generating new images. This implies that after going through the
generator, certain features of the input picture should be saved. The adversarial loss, on the other

54
hand, cannot ensure that the learnt function will translate the input to the intended output.
Additional loss functions must be employed to limit the space of potential mapping functions
even more.

Loss of cycle consistency.

The generator network may decrease the mapping function space karat by ensuring that the
function is cycle-consistent. This loss was suggested for unpaired dataset training. However, it
has shown positive results when used in a paired dataset arrangement. This loss was shown to be
useful in color or texture changes, such as style transfer or picture enhancement. However, when
the transformation involves large changes in the picture, it performs badly. The original
CycleGAN had two generators, one learning a transformation and the other learning the reverse
transformation.

3.5.Research Tools
The research tools utilized for the experiments are determined by the categories below. It might
be hardware or software that is used for both research implementation and documentation
purpose.

3.5.1. Software Tools

The prototype is developed with the programming language Python using TensorFlow and Keras
for neural network construction, training and computation. Software Tools are used for
developing software, debugging, visualizing findings, and documenting the research process in
preparation for reporting. The following lists determines the research tools used for the
experiments.

TensorFlow and Keras

TensorFlow is an interface for building machine learning techniques including Deep learning
developed by Google. A TensorFlow computation is defined by a graph of data flows. A graph
contains a set\s of nodes where each node indicate a mathematical process. Not all but certain
type of nodes are permitted to retain and change their permanent state. Data that flows between
nodes are represented by edges. Data are represented as tensors, unrestricted multidimensional
arrays. Since the graph merely represents computations, one need to lunch the graph in a

55
distributed session in order to calculate anything. A computational graph is built using one of the
available programming languages like Python.

Keras is a deep learning library capable of running on top of libraries such as TensorFlow to
offering a high-level neural networks API for fast and easy prototyping. It is written in Python.

Python and Jupiter Notebook

Python is a programming language that helps you work quickly and interact systems more
efficiently. Python programming used to build and illustrate this thesis needs many of the drivers
needed to setup and bundle to install some devices with the computer. A Jupiter Notebook is an
intuitive electronic application that designs, load python Programming interface, and compose
python code.

Azure Virtual Machine

Azure Virtual Machine is one of many kinds of on-demand, scalable computing resources that
Azure provides, and used in this research for testing different models that are used for other
different purpose and datasets.

Google Collab

Google Collaboratory Laboratory enables anyone to write and run arbitrary python code via the
web, and is highly suitable to machine learning, data processing and training and has free access
to computing resources, such as GPUs.

Microsoft offices Tools

Microsoft Office pieces include the word processor (Word), the spreadsheet (Excel) and the
visual presentation tool (PowerPoint.) is used in this research for document processing and
reporting also, for to label pictures with their file name together with its matching class label
number.

Citation-machine and cite-this-for-me

Both are used to help you quickly prepare an entire bibliography or reference list in simple steps,
and format with the intended format, for this research the reference format is IEEE.

56
 3.5.2. Hardware Tools
Hardware: Laptop PC

 RAM: 8GB
 Window 10 Intel(R) Core(TM) i7-6500U CPU @ 2.50GHz, 2601 MHz, 2 Core(s), 4
Logical Processor(s) HP laptop, Disk: 1TB
 For taking design and dress pictures GALAXY J7 Phone camera

57
 3.6.Evaluation Metrics
In supervised learning using a classifier: There are evaluations we may pass on which contain
labels on how our images should be categorized. And based on these labels, we can assess
whether our model is correct or erroneous. However, using a GAN, we feed in some random
noise to it and we obtain this fake image. Nevertheless, there is no definitive way of knowing
how realistic these produced pictures are. We do not know the precise pixels it has intended to
produce, we cannot tell. Therefore, since there is that, there is no clear objective for what pixels
we are meant to produce given this noise vector coming in. So, this approach is more like an
expert artist learning how to create wonders. Additionally, the discriminator in this GAN, which
distinguishes genuine vs false, does not ever achieve perfection. In addition, frequently overfits
to distinguishing genuine from false pictures for its specific generator. But then, how do we
evaluate GANs?

Well, we may start by first establishing the desirable criteria we want from these characteristics
we would want. And one major characteristic, one primary goal is definitely fidelity or the
quality of our produced pictures and how genuine they appear. A decent generator also generates
a good range of pictures; it would not be very helpful to only produce this one image here.
Therefore, the second characteristic we usually worry about is variety. What is the range, what is
a diversity of pictures this generator is able to produce? Is it able to generate the variety that is
inherent in the training data set or the intended classroom modeling? Therefore, in summation, it
is difficult to assess again since there is no universal discriminator. Giving a ground truth that
would allow fair comparisons among GANs. In addition, to assess a GAN, we want to examine
fidelity or the quality of pictures. As well as the diversity or variety of the pictures coming out of
our GAN.

3.6.1. Fréchet Inception Distance (FID)

Fréchet inception distance, or FID, is the most common measure for assessing the feature
distance between actual and produced images. Evaluation is very much an open topic in
generative model’s research. So Fréchet distance named for the mathematician, Maurice Fréchet
is a distance metric that is used to quantify the distance between curves and may be extended to
compare distributions as well. You can compute the Fréchet distance between distributions.

58
There are numerous distributions for which the Fréchet distance between two distributions is
analytically solved. FID looks at the statistics of the actual multivariate normal distribution and
the statistics of the fake multivariate normal distribution, where the statistics are simply mean
and covariance matrices, and determines a how far away those data are from one other. Where
the closer those statistics are to one other, the closer the false embedding mimic the actual
embedding. If the fake is near to the real, meaning your model is producing fake outputs close to
the real, the smaller this difference this number is. A smaller number really indicates that
characteristics in the reals and fakes are more comparable to one other thus that means lower the
FID, the better. Usually, the closer you are to zero, the better. Because at zero, your fakes are
indistinguishable from the reals based on the statistics derived from numerous of their
embedding. The issue is that current assessment measures are not precisely the best indication of
success today. Nevertheless, FID is still one of the most common methods to assess your GAN
since it is simple to build and utilize.

Its computation is based on an Inception model. However, it utilizes the statistics of an

intermediate layer, which is modeled by a multivariate Gaussian distribution N (µ; Σ) (µ is mean
vector, Σ is covariance matrix). For real samples, the feature distribution is denoted as N (µx;
Σx). Correspondingly, the feature distribution of generated samples is N (µx̂ ; Σx̂ ).

FID is the Fréchet distance between them given by

1
2 2
∥ μ x −μx^ ∥ +Tr (Σ x + Σ ^x −2(Σ x Σ x^ ))
2

where Tr sums up all the diagonal elements.

3.6.2. Inception Score

Inception score is another method to measure the distance between reals and fakes using the
Inception v3 model. This technique was created before FID and used to be extensively utilized
but now has been substantially superseded by FID and is by far more relevant to conditional
generation. the inception score also utilizes Inception v3 classification model. However, unlike
FID, you utilize the classifier as is not as a feature extractor. You leave the classifier intact and
don't utilize any of those intermediate results. First, you enter a created image and check what the
inception v3 classifier thinks it is categorized as. For example, maybe this picture seems like its
0.5 dog, 0.4 cat, and 0.1 bird. Intuitively a high rating on one class indicates that this picture

59
obviously resembles one class over the others. That is, it is arguably high quality. You would
like this diversity probability distribution across all the various classes to be well spaced out and
not spiky on a single class. More technically, the first measure which looks at the distribution of
classes given an image, is measuring the distribution p of y given x, where given your picture x,
what is the distribution across those classes given that classifier. What this is attempting to
mimic is that there are obvious things being selected out in the produced pictures by your
classifier. This is characterized as having low entropy since the probability distribution is
concentrated around those chosen spots and not spread all about. Entropy in a sense is
unpredictability, and this fidelity measurement correlates to low entropy. Now on the diversity
front, you want many of classes to be produced among your numerous samples. The distribution
of classes among your samples should have high entropy, meaning it is not focused on a single or
a few of classes, since it implies you are just a producing one kind of thing and nothing else. This
distribution is also known as the marginal label distribution or p of y. This is the distribution
across all of your labels, all labels over your whole dataset or across a big sample. This may be
confusing because you have high down here and you have low up here. However, just bear in
mind that these two are meant to be extremely distinct, that is what is important here. With this
notion of fidelity being low entropy and of diversity being high entropy, you could combine their
measurements into a score, which is convenient because then you can score your model and say
I'm getting better or I'm getting worse by getting a distance between that diversity are high-value
minus fidelity, a low value.

Generally, the calculation is based on a pretrained Inception V3 model, which was initially
suggested for image classification and produces a list of probabilities p (y|x) € [0; 1] N for each
input picture x. The number of labels in the dataset is N, and y is the set of class labels.

Inception Score is equal to exp(ExDKL(p(y|x) ||k p(y)) )

where x is a GAN-generated image (i.e., x = G(z)), p(y) = ∫p(y|x = G(z)), dz is the marginal label
distribution. The conditional label distribution p(y|x) should be narrow, indicating that the
picture is crisp and includes a discrete object. The marginal label distribution p(y) should, on the
other hand, be uniform, suggesting that the pictures come from all classes and are diverse. The
KL divergence between the two distributions will be high if both of them are fulfilled. As a
result, a higher IS indicates that the GAN can perform better.

60
 3.6.3. Nearest Neighbors
The user assessment evaluation is conducted by showing both the real and generated images in
the proposed method and take the inputs from the Professional Designers and Crowdsourcing.

Chapter 4. Design And Experiment

This chapter deals with the design of the proposed model and associated experimental settings.
Particularly, the architecture of the proposed model and explanations, how information can be
retrieved and covers the development of the prototype using the dataset created to generate new
design variations in the proposed model and other pretrained models.

4.1.Proposed Model and Model Description

The obtained data would first be preprocessed and the data will be filtered using various
techniques during the preprocessing task in order to obtain better data. The preprocessing task
includes a variety of sub-tasks such as image form conversion for ease of computation, image
enhancement for visual ease, noise reduction for process accuracy, image scaling and resizing or
shrinking for complexity minimization, Aggregation, Augmentation, clear background and other
essential techniques that are needed to improve data quality. Furthermore, preprocessing tasks
may have an impact on a classifier's overall output since they can expose unknown or unclear
details in a file. Therefore, after that data processing labeling will be done based on different
conditionals; such us color, shape, size, texture, pattern, texts, stripped and others related
conditions. This will be done by using either Outsourcing , domain experts and by Transfer
learning from others similar research datasets.

This study’s choices of model from various machines learning aspects is Generative Adversarial
Networks that can create fake image but look realistic images with the utilizing of two neural
networks namely “Generator” and “Discriminator” in which they act against one another, that's
why the model have the name "adversarial." Both neural networks have specific roles. One is a
convolution neural network that plays the role of a discriminator/classifier of the generated
images whether they are real or fake, and the other a reverse convolution neural network that
plays the role of a generator which is responsible for generating fake images.

Finally, this study will train a new GAN (Generative Adversarial Network) model by combining
different generative adversarial networks model to generate a novel fashion image with the best

61
quality, good aesthetics and beauty, also based on different user’s fashion design conditionals.
Then the model will be evaluated for validation purpose.

The overall frame work architecture is look like the following picture.

Figure 16: The Proposed Model framework

As it shows in the above image the overall steps to implement the proposed model are the
following:

 Collect and gather both design images and design sketches

 Separate them in to two datasets for both type of image dataset
 Train the first StyleGAN using Transfer Learning, produce enough image data
 Train the second StyleGAN using Transfer Learning, generate enough image data
 Final from both the generated images train CycleGAN for image-to-image translation and
produce the desired novel fashion designs

62
  After that evaluate all three-model results using Fréchet inception distance and Nearest
Neighbors.
 Finally conclude the results and make it applicable for real time scenarios.

4.2. Dataset Preprocessing

Data collection is the most important step of any research endeavor. The dataset would have
been after conducting relevant study in domain as if dataset is not appropriate for the project, and
the intended objective of the research may not be accomplished.

Data augmentation is a commonly used machine learning method that aims to increase the
amount of the training dataset in order to improve the generalization capabilities of a machine-
learning model. Data augmentation is the process of creating new data instances from existing
data examples in the training dataset. In object recognition, data augmentation techniques must
maintain the contents of the images so that the characteristics that are helpful for recognizing an
item are present in the produced images.

Some techniques in computer vision stand out primarily of their popularity:

Flipping: Reversing a picture across one of its axes is known as flipping an image (horizontal or
vertical).

Random rotation: This method rotates the input picture a random number of degrees, as the
name implies. Rotation may also be done in a set number of degrees (e.g., 90 degrees), but at
random intervals. This method is only effective when the model's target object's label does not
change as it rotates. Rotating an image of an eight may, for example, result in an image with the
sign, which is not the same symbol.

Random cropping: This method chooses a rectangular portion of the supplied picture at
random.

Rescaling: This is the process of altering an image's size while maintaining the whole of its
information. Objects in the rescaled picture usually do not have the same proportions as those in
the original image.

63
Adding random noise: This method modifies the look of the input image by adding some
random (typically gaussian) noise to each of the pixels.

Techniques for advanced users: Some innovative techniques, such as GAN data augmentation,
use a generative adversarial network to generate entirely new pictures. They allow for a
considerably greater range of pictures to be generated than conventional data augmentation
techniques.

The dataset shape may be described by shape of a picture as each cell consists of a pixel value.
Therefore, the total number of rows reflects the number of images and total number of pixel
values represents number of columns. The structure of training images before preprocessing:

Images exist in different dimensions (form, size, orientation, background etc.) making it
inappropriate for computer vision applications, since it would be difficult for the neural network
to extract features. To prevent this, we need to orient (pre-process) the data (images) properly
before it is provided as input to the network. Open Source Computer Vision (OpenCV) and PIL
python library is utilized in the project for this purpose, which includes approximately 2500
algorithms to operate on images.

Normalization and Background Removing: ImageNet’s standard deviations and mean values
for each of the three RGB channels were used to normalize the pixels in the images. This is a
frequent technique in image classification research since the normalization values for different
problems are often comparable, and calculating the mean and standard deviation of each channel
in a new dataset may be costly.

Reshape and Augmentation: Parallel Flipping, Shear Range, Width Shift Range, Height Shift
Range, Zoom Range, and other augmentation parameters are used to create the desired images.
Random Horizontal Flip This function, as the name implies, takes an input picture and, with a
specified probability p, outputs the identical image reversed horizontally. As stated previously in
this thesis, this technique was developed as a data augmentation policy that maintains the
aesthetic value of the pictures. Finally, since the dataset is expanded by utilizing various
methods, a sufficient number of pictures are produced.

64
 4.3.Experimental Setup and Implementation
The primary task is to ensure that each image in the data collection has the same format in terms
of dimension, extension, and background. Any irregular images will be automatically
transformed to the required image and saved to the dataset. The training is done on Google Colab
Pro since GPU is required and GAN will not train effectively only with a CPU. We must mount
the dataset from the disk to the cloud. The training needs square pictures, and images with a size
of 1024 x 1024 are required for extremely high quality. However, this study used a resolution of
512*512, and the next step will be to resize all of the pictures to this resolution and use PIL,
which is a Python library to transform an RGBA image to an RGB image. After opening a
picture, use Image.new(mode, size, color) to generate a new RGB image with mode set to
"RGB", size set to the desired image size, and color set to (255, 255, 255). Using Image.split and
Image.paste (picture, box, and mask) with image as the original image and mask the alpha
channel. Image.split () extracts a tuple of separate image bands from an image (red, green, blue,
alpha). Index into three to obtain the alpha channel. Then call Image.save (filename, format,
quality) to save the improved image with the filename as the filename, the format override
format, and the new image's quality.

# resize all the images to same size

import os
from tqdm import tqdm
import cv2
from PIL import Image
from resizeimage import resizeimage
path = '/content/drive/MyDrive/DataSet/samesized/samesized/'
for filename in tqdm(os.listdir(path),desc ='reading images ...'):
image = Image.open(path+filename)
print(image)
rgba_image = image
rgba_image.load()
background = Image.new("RGB", rgba_image.size, (255, 255, 255))
background.paste(rgba_image, mask = rgba_image.split()[3])
image =background
image = image.resize((512,512))
image.save(path+filename, "JPEG", quality=100)

The sample Image after the Data preprocessing take place is shown below:

65
 Figure 17: Dataset Random capturing's

During the model training process, each input picture is scaled to 512x512 pixels in size. The
network, like the StyleGAN, is trained from scratch. Once the custom dataset has been
transferred to Colab and resized, by using the following commands to convert the custom images
to tfrecords. This is a StyleGAN requirement, therefore this step is required for training our
Model.

!python /content/stylegan/dataset_tool.py create_from_images /content/

stylegan/datasets/custom-dataset /content/drive/MyDrive/samesized
Once the tfrecords are successfully created, you should be to view them under like:

These tfrecords correspond to 4x4 , 8x8 ,16x16,

32x32, 64x64, 128x128, 256x256 and 512x512
resolution images (baseline progressive)
respectively.

Figure 18: Dataset tfRecords result

66
Now go to our training folder and open the train.py file with our generated tfrecords custom
HabeshaDressSet. Which the implementation looks like

desc += '-/content/stylegan/datasets/custom-dataset'; dataset = EasyDi

ct(tfrecord_dir='/content/stylegan/datasets/custom-dataset');
train.mirror_augment = True
And then we set the number of GPU’s. since this research is conducted on Colab pro plus it only
allows 1 GPU, therefore we set our training with this GPU.

desc += '-1gpu'; submit_config.num_gpus = 1; sched.minibatch_base = 4; sch

ed.minibatch_dict = {4: 128, 8: 128, 16: 128, 32: 64, 64: 32, 128: 16, 256
: 8, 512: 4}
Then this study use the number of epochs to 12000 because of the longest time that Colab can
allow us and it is efficient for our resolution type. Training will stop after this much iterations.
Then we set the inception pkl file form StyleGAN since it shows effective result in generation of
quality celebrity faces and different datasets.

Also, this research uses transfer learning for inception features from StyleGAN. Therefore the
inception_v3_features.pkl file from StyleGAN GitHub to our training folders. And insert in our
Fréchet distance file.

We are now all set to go for training our own Model. After the training Once, our training is
over, there will a directory created that contain results of the training which include a lot of
network-snapshot pickle files created. we need to get the latest .pkl file and use the weights of
that file.

4.4.Implementation of Model Evaluation

The implementation of Fréchet Inception Distance is the following.

# Calculate FID.
m = np.square(mu_fake - mu_real).sum()
s, _ = scipy.linalg.sqrtm(np.dot(sigma_fake, sigma_real), disp=False
) # pylint: disable=no-member
dist = m + np.trace(sigma_fake + sigma_real - 2*s)
self._report_result(np.real(dist))

67
Chapter 5. Results and Discussion
After successfully training the GAN model, it is now able to generate new samples that look
similar to the training dataset. In this section, the experimental results of the model are shown to
understand the usefulness of the trained generator in fashion design process. Two tests have been
conducted in the identical hardware and software, hyperparameters, and dataset setup. each
experiment suggested insights by employing transfer learning and introducing an additional
condition style based generative adversarial network.

5.1.Gathering of Image Data

The picture data collection technique described in the preceding chapter, along with some human
cherry selecting, resulted in a dataset of over 35,000 fashion-related photographs. Approximately
2/5 of the photos given by the algorithm were included in the dataset, while the remaining 3/5
were eliminated during the human selection process. The manual selection was done with the
goal of improving our chances of training a generative model by restricting the data to the
required viewpoint (front view) and removing distracting elements that obscured the dress
Image. If we estimate that manually downloading and cropping three photographs takes roughly
two minutes, manually creating a dataset of the same size would take many days. That does not,
however, involve network data cleaning and preparation, which the chosen system can do
automatically. The chosen system and tools rapidly delete files that are too tiny, corrupt, or of an
undesirable file type, as well as shrinking photos that are too big and removing duplicates from
the collection. Manually doing the same operation would be very laborious and time-consuming.
Several variables might have contributed to a reduction in the number of photographs that had to
be eliminated during the manual selection process.

In the TensorFlow Item Detection areas, there are additional models that output masks of the
observed object rather than merely a bounding box. It would be able to eliminate the background
of the photographs and just maintain the required item using a masking model, resulting in a
much cleaner dataset with no fluctuating background. This was investigated, however we opted
against using them since they need considerably more extensive prelabeled data than bounding
box models. However, it is something worth considering in the future.

68
5.2.Generated Samples with different Iterations

(a)Initial Iterations (B) After 1000 Iterations

(C) After 2500) (D) After 4000 Iterations

69
(E) After 5500 Iterations (F) After 6800 Iterations

(G) After 7100 Iterations

Figure 19: List of Images based on Training Iterations.

70
5.3.Generated Image from StyleGAN
Table 7: Generated Images Using StyleGAN Model

71
 5.3.1. Failed examples from StyleGAN
These bad examples occurred due to various reasons. The first would be irregular shapes and unidentified shapes along with a
different class in the dataset distribution.
Table 8: failure examples

72
5.4.Generated Image from StyleGAN2 with adaptive discriminator augmentation (ADA)
Table 9: Generated Image from StyleGAN2 with adaptive discriminator augmentation (ADA)

73
 5.5.Evaluation Results
5.5.1. Quantitative results
Table 10 : List of Networks snapshot with their respective FID result of StyleGAN

Network Snapshots FID Results

network-snapshot-000140 fid50k 391.0658
network-snapshot-001283 fid50k 301.6796
network-snapshot-002364 fid50k 280.7590
network-snapshot-003285 fid50k 262.8640
network-snapshot-004085 fid50k 203.2286
network-snapshot-004705 fid50k 64.0024
network-snapshot-005306 fid50k 40.899
network-snapshot-005726 fid50k 35.1669
network-snapshot-006900 fid50k 18.6832
network-snapshot-007170 fid50k 20.9514

Table 11: List of Networks snapshot with their respective FID of StyleGAN2 with adaptive
discriminator augmentation (ADA)

network-snapshot-000000 fid50k_full 305.7188

network-snapshot-000080 fid50k_full 283.2060
network-snapshot-000160 fid50k_full 91.4158
network-snapshot-000240 fid50k_full 42.5183
network-snapshot-000320 fid50k_full 26.8195
network-snapshot-000400 fid50k_full 18.2750
network-snapshot-000480 fid50k_full 14.5784
network-snapshot-000560 fid50k_full 13.9135
network-snapshot-000640 fid50k_full 13.4357
network-snapshot-000720 fid50k_full 12.1136
network-snapshot-000800 fid50k_full 14.4513
network-snapshot-000840 fid50k_full 12.1395
network-snapshot-000880 fid50k_full 11.1006

5.5.2. Nearest Neighbors results

The purpose of our user research or nearest neighbors was to assess if the provided model
prototype findings promotes creative thinking, practical usability, and ideation for fashion
design. The key topics highlighted during the investigation were emotional reaction, control and
usability and descriptive analysis. An investigation of the visual consequences is next
investigated. The purpose of the qualitative assessment was to offer further support to the
reflective component of the creative process. Outcomes from the assessment focused on

74
qualitative interpretations of interactions between participants via a focus group session that
comprised a questionnaire and usability-testing module. This was aimed to give a significant
primary source of data for analysis directly relevant to the study topics, via photos and completed
questionnaires. Analysis of the data was conducted utilizing thematic techniques with concerns
towards user - centered design processes. The goal of the study was to give greater insight into
overall results and individual reactions of participants by recording the attitudes and emotional
responses to the technology, while finding hidden topics that reveal underlying concepts,
patterns, and assumptions.

Nearest Neighbors with random people

Twenty individuals in all were part of the experience, representing a broad variety of random
peoples with various backgrounds mostly in painting. Ages varied from 19 to 35 with twelve
male and eight females. Many have not considerable knowledge and expertise within the visual
arts but they had a little understanding via education, career and individual creative activity. A
big component of the descriptive analysis was describing the picture quality of outputs. This
usually had a negative tone as participants objected to digital artefacts and poor quality of the
photographs. Connecting the visual outputs from individual findings to the questionnaire also
gave insights into the connection between user attitude and behavior. In this assessment,
consumers got an opportunity to evaluate the more realistic photographs from two options. Based
on the user research assessment some people describe the genuine as fake and the false as real.

20
18 19
16
16
14
12 13
10
8
8
6
6
4
2
2
0
Category 1

I like the Idea It feels like real (Quality) I don't get it

The design created by Computer It doesn't show creativity The idea is Promising

Figure 20: Nearest Neighbors assessment from random people.

75
This study also gathered user comments on the outcomes, and the results suggest that more than
half of the participants can tell the difference between genuine and false images based on our
created images. This test only looks at image quality. Each test case contains two images: (1) the
original image and (2) the edited image. Users provide values ranging from 0 to 10 to both
images depending on image quality.

Nearest Neighbors with Professional Fashion designers

The researcher conducted a user study with twelve fashion design professionals (8 females and 4
males) who are currently working in the fashion design industry and teaching in fashion design
schools the graduate class of Fashion design university students are included. Note that these
participants were newly recruited for this GAN model evaluation study.

During the interview, the participants get some knowledge on artificial generated image and how
it can help them to produce new ideas. This helped designers determine and expand the range of
styles in their repertoire. Participants mentioned that they were able to expand the scope of
concepts or increase the number of concepts in a specific style by means of this new sector of AI
generated fashion design. In addition, they believe that in this model style, there was a design
that they had not usually thought of, but it was a recommendation in an understandable range,
which expanded the range of the style they were thinking of.

The researcher did a user study with twelve fashion design professionals (8 females and 4 males)
who are now working in the fashion design business and teaching in fashion design schools the
graduation class of Fashion design university students are included. Note that these individuals
were freshly recruited for this GAN model assessment research.

During the interview, the participants obtain some understanding about artificial produced
picture and how it might aid them to develop fresh ideas. This helps designers select and increase
the spectrum of styles in their repertory. Participants noted that they were able to broaden the
breadth of ideas or enhance the quantity of concepts in a certain style by means of this new
sector of AI produced fashion design. And they feel that in this model style, there was a design
that they hadn’t generally thought of, but it was a suggestion in an intelligible range, which
enlarged the range of the style they were thinking about.

76
 5.6.Discussion
In this research, we have studied how AI systems, especially in the area of deep learning, might
be implemented into the design management process to make processes creative, quicker, easier
or more effective. The intention was that this study would lead to broad recommendations and
enable us to construct a framework to motivate, support and guide organizations and designers
interested in taking a first step in applying artificial intelligence to generate innovative designs.

The work process started with a fairly broad scope to become acquainted with the subject of deep
learning. When adequate information had been gathered, through reading various papers and
study results, the scope was reduced to concentrate on my passion and Fashion industry. The
implementation stage of presented concept was employing transfer learning using StyleGAN
with TensorFlow to create and train neural networks in Python. The decision of utilizing these
Model and libraries over others was determined early based on primarily it is applicability for
this research and acceptance, community sizes and for being open source. They are also
extremely well documented with various examples and tutorials, which have been a helpful
resource along the road.

When it comes to the custom dataset, the model had some challenges to create synthetic pictures
equal to the real data. Several explanations are behind the issues. First, several images inside the
collection contain some noise in terms of other things besides Habesha garment, are included in
the images. If the dataset is inspected, one may discover things such as a two dress in one
picture, logo and trademarks, store location, Traditional materials and many other objects.
Moreover, some photos have one dress while others might contain numerous outfits. All these
characteristics make it harder for the Model to gather the features of the classic dresses and
designs. Secondly, the size of the dataset also plays a factor.

For the habesha dresses images dataset there are more than 13000 samples. The final argument
concerns the capacity of the StyleGAN assuming that it is not already adequate. The model may
require some further fine-tuning in order to become better however due to lack of time the
researcher did not go on with it. Comparing the findings from the two different sizes employed,
training a model on bigger photographs proved to boost the visual quality of the images. One
explanation might be because details are overlooked in images of reduced size.

77
A challenging problem that affects the outcomes from an experiment is that they must be both
reliable and accurate. Reliability refers to the repetition of results that is when a phenomenon is
tested repeatedly, the conclusion must be relatively comparable. Additionally, objective metrics
are more accurate than subjective assessments. The researcher assure that the findings from the
experiments are reliable because every experiment was done at least triple, and every time an
experiment was run extremely comparable outcomes in quality and variation were produced.
Anyone who conducts the tests that have been done in this paper using the same technology
(hardware and software), datasets and setups with identical period, which is 72 hours, should
achieve comparable conclusion.

The suggested model revealed that it can create it is possible to develop designs that are
comparable to human designs. However, human creations are not only include technical details,
but also demonstrate considerably wider variation. This is for multiple causes. First, people may
apply their much greater past knowledge for the Fashion design work, whereas machine can only
gain restricted information from the provided data. Second, a human designer is able to assess
the design job fairly and logically, which may assist enhance their designs. By contrast, without a
proper evaluating function, machine can only conduct pilot" creation". Integrating greater past
information into the deep networks, as well as building realistic assessing functions, are two
significant paths of future study for producing more creative and credible outputs of dress
design.

Since this thesis attempts to encourage and motivate the usage of artificial intelligence it is
important addressing the probable long-term repercussions. The automation enabled by artificial
intelligence, driven by deep learning, might certainly lead to the termination of some
conventional occupations. This is something that goes hand in hand with technology
improvements. Since the implications of artificial intelligence on diverse professions is
unpredictable, the transition needs a high degree of flexibility and preparedness for multiple
result scenarios in order to be seamless.

78
Chapter 6. Conclusions and Recommendations
6.1.Conclusions
The major purpose of this study is to look at how generative adversarial networks (GAN) may be
used to allow automated fashion picture production and how such technologies might help
professional fashion designers with their creative designs.

As a result of the current advancement of GANs in the fashion sector, we were able to generate
some new fashion images with improved quality, create virtual cloth generation consisting of
new designs and patterns, study the distribution of fashion images, and develop new clothing and
accessories that maximize users' interests. GAN may be used in a variety of Fashion suggestion
scenarios, such as discovering the world of new fashion goods, changing existing items, or
developing items tailored to a person and assisting in the creation of new products. Another key
item to note is that, like with practically every ML technique, data pretreatment is critical for
training GAN's model performance. The findings clearly showed that the chosen could readily
create high quality samples of other datasets; however, the bespoke Habesha clothing dataset
images had issues because the dataset was not well preprocessed and had some noise.

6.2.Limitations and Recommendations

Latest studies on Fashion image generation using generative adversarial networks showed some
development. Even though GAN has proven large and beneficial results in the past, it is still a
problematic issue for some reason. The gap shown by earlier studies are the quality of the
generated images is inadequate and incapable of providing the control of really well styles and
not take into account some conditionals to be shown in generated images such us, visual artifacts
patterns, caption, a huge set of colors and diversity in their datasets, which is generative
adversarial network is majorly reliant on the number of training images. With very low input, the
system will not be able to understand detail of the desired picture, so the size of each domain has
to be quite huge. In addition, not including the viewpoints of experienced fashion designers has a
lot of influence in refining the results.

One drawback of this research offered methodologies including other deep learning methods is
that a considerable quantity of data is needed for training. This is especially a problem for
traditional fashion image processing, in which traditional clothing pictures are costly and rare to

79
get. In order to strengthen the actual usage of the model, the researcher want to undertake further
experiments by utilizing fewer training samples and further develop strategies and data
augmentation to improve data productivity.

Generally, the limitations of this thesis are considering only Ethiopian Habesha Dresses, Due to
Resources and Time constraints both training time is for 72hrs, currently it is only helpful for
Designers to input ideas for them and show some styles that they were not thought of before, and
Not supporting Texts Input and patterns from Customers.

In the future, this opens up a potential line of study in applying GAN’s architecture for
Generating and recommending fashion design. The next work need to investigate boosting the
quality and resolution of the produced pictures and giving control of fine-grained styles, the
same principles may be extended other visual data besides fashion photographs, or even to non-
visual forms of material. Evaluation of GAN’s performance and preserve the realism of human
postures and faces will also be an essential subject to research.

The present approach creates unique and believable naturally pictures in a random way. A
logical issue is if we can enhance this automated design, process to be goal-oriented like human
designers would do. For example, the design outcomes may be repeatedly developed according
to specific criteria, where methods such as reinforcement learning can be utilized. However, this
needs excellent quantitative measures on originality and utility of the created products, which
remains a tough unresolved topic. Evaluation of the produced results technique are likewise an
open research topic.

Therefore, the following points are planned to be implement in the future;

 Fully working Intelligent Web system that lets customers to design their Innovative
Fashion Dresses in their desires.

 Combine with text inputs from customers

 Boosting the quality and resolution of the produced pictures and giving control of fine-
grained styles

 Enhance this automated design process to be goal-oriented like human designers

would do. i.e. repeatedly developed according to specific criteria

80
References

[1] Imran Amed, Anita Balchandani, Jakob Ekeløf Jensen, Achim Berg, Felix Rölkens, "The
State of Fashion 2021," McKinsey and Company, 2021.

[2] N. i. B. D. v. d. Pols, "BUSINESS OPPORTUNITY REPORT ETHIOPIA TEXTILE &

APPAREL INDUSTRY," 2021.

[3] U. P. Amity University, "Fashion & Textile Industry 4.0 -Opportunities & Challenges for
Education 4.0," in Fashion Apparel and Textile (NCFAT’20), India, 2020.

[4] A. W. Dessalgn, "GENERATIVE ADVERSARIAL NETWORK-BASED VISUAL-

AWARE INTERACTIVE FASHION DESIGN FRAMEWORK," vol. 1, p. 24, oct, 2020.

[5] Christie's, "Is artificial intelligence set to become art’s next medium," Christie's, 12 12
2018. [Online]. Available: https://www.christies.com/features/A-collaboration-between-
two-artists-one-human-one-a-machine-9332-1.aspx. [Accessed 24 march 2021].

[6] D. Flaherty, "Artistic approaches to machine learning," vol. 1, p. 13, 2020.

[7] Rajkishore Nayak, Rajiv Padhye, "Artificial intelligence and its application in the apparel
industry," Automation in Garment Manufacturing, pp. 109-138, 2018.

[8] "AI in Fashion: An Extensive Guide To All Applications For Retail," Intelistyle, 03
December 2020. [Online]. Available: https://www.intelistyle.com/ai-fashion-retail-
applications/. [Accessed 04 April 2021].

[9] M. S. Kim, "AI as a Designer?," Medium, 08 August 2020. [Online]. Available:

https://medium.com/sfc-interaction-design-class-2020-spring/ai-as-a-designer-
3fce8db9451e. [Accessed 04 April 2021].

[10] "Artificial intelligence in the fashion industry," Research Features, 04 March 2021.
[Online]. Available: https://researchfeatures.com/artificial-intelligence-in-fashion-
industry/. [Accessed 04 April 2021].

[11] "How to Use Artificial Intelligence (Ai) in Online Fashion Retail in 2021," Exponea, 12
January 2021. [Online]. Available: https://exponea.com/blog/impact-artificial-intelligence-
online-fashion-retail/. [Accessed 04 April 2021].

[12] Marie Korneliusson, John Martinsson, Olof Mogren, "Generative Modelling of Semantic
Segmentation Data in the Fashion Domain," 2019 IEEE/CVF International Conference on
Computer Vision Workshop (ICCVW), 2019.

[13] Othman Sbai, Mohamed Elhoseiny, Antoine Bordes, Yann LeCun, Camille Couprie,
"DesIGN: Design Inspiration from Generative Networks," Lecture Notes in Computer

81
 Science, pp. 37-44, 2019.

[14] Katrien Laenen, Susana Zoghbi, Marie-Francine Moens, "Web Search of Fashion Items
with Multimodal Querying," Proceedings of the Eleventh ACM International Conference
on Web Search and Data Mining, vol. 1, 02-2018.

[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,
Sherjil Ozair, Aaron Courville, and Yoshua Bengio, "Generative adversarial nets.," In
Proceedings of the 27th International Conference on Neural Information Processing
Systems, vol. 2, 2014.

[16] Donggeun Yoo, Namil Kim, Sunggyun Park, Anthony S. Paek, In So Kweon, "Pixel-
Level Domain Transfer," European Conf. on Computer Vision, 2016.

[17] B. Roberts-Islam, "Why Fashion Needs More Imagination When It Comes To Using
Artificial Intelligence," Forbes, 21 September 2020. [Online]. Available:
https://www.forbes.com/sites/brookerobertsislam/2020/09/21/why-fashion-needs-more-
imagination-when-it-comes-to-using-artificial-intelligence/?sh=3b97ad7e3f63. [Accessed
04 April 2021].

[18] Radford A, Metz L and Chintala S., "Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks," 2015.

[19] Tero Karras, Tero Karras, Tero Karras, jaakko Lehtinen, "PROGRESSIVE GROWING
OF GANS FOR IMPROVED QUALITY, STABILITY, AND VARIATION," 2018.

[20] Tero Karras, Samuli Laine, Timo Aila, "A Style-Based Generator Architecture for
Generative Adversarial Networks," Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, 2019.

[21] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, "Image-to-image translation with conditional
adversarial networks," in IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2017.

[22] Kristen Vaccaro, Tanvi Agarwalla, Sunaya Shivakumar, and Ranjitha Kumar., "Designing
the Future of Personal Fashion. In Proceedings of the 2018 CHI Conference on Human
Factors in Computing Systems (CHI '18).," in Association for Computing Machinery,,
New York NY, USA, , 2018.

[23] D. F, "Can a tool generate multiple fashion designs in a few minutes? Yes, with the help
of machine learning.," Medium, 19 March 2021a. [Online]. Available:
https://deboraf.medium.com/can-a-tool-generate-multiple-fashion-designs-in-a-few-
minutes-76ae26d2af92. [Accessed 22 march 2021].

[24] Xintong Han, Zuxuan Wu, Phoenix X. Huang, Xiao Zhang, Menglong Zhu, Yuan Li,
Yang Zhao, Larry S. Davis, "Automatic Spatially-Aware Fashion Concept Discovery,"

82
 2017 IEEE International Conference on Computer Vision (ICCV), pp. 4-12, 2017.

[25] W.-C. Kang, C. Fang, Z. Wang, and J. McAuley, "“Visually-aware fashion

recommendation and design with generative image models," IEEE International
Conference on Data Mining (ICDM),, 2017 .

[26] Othman Sbai, Mohamed Elhoseiny, Antoine Bordes, Yann LeCun, Camille Couprie,
"DeSIGN: Design Inspiration from Generative Networks," Lecture Notes in Computer
Science, Cham: Springer International Publishing, 2019.

[27] "CVPR 2020 Workshop," Seattle, Washington,, June 2020.

[28] "Fashion Dictionary," WWD, March 02, 2015. [Online]. Available:

https://wwd.com/fashion-dictionary/. [Accessed August 23, 2021].

[29] "What is fashion design?," Fibre2Fashion, april 23, 2021. [Online]. Available:
https://www.fibre2fashion.com/industry-article/2860/what-is-fashion-
design#utm_source=f2f&utm_medium=content&utm_campaign=interlinking. [Accessed
August 2, 2021].

[30] "Fashion design," Wikipedia, August 22, 2021. [Online]. Available:

https://en.wikipedia.org/wiki/Fashion_design . [Accessed August 23, 2021].

[31] n.d., "Traditional Dress," LoveToKnow, [Online]. Available: https://fashion-

history.lovetoknow.com/clothing-types-styles/traditional-dress. [Accessed August 24,
2021].

[32] Ayele, Gezahegn, Moorman, Lisa, Zhang, Xiaobo, "Infrastructure and Cluster
Development: A Case Study of Handloom Weavers in Rural Ethiopia," The Journal of
Development Studies, 2011/12/01.

[33] N. TV, "Facebook," NahooTv, [Online]. Available:

https://www.facebook.com/Nahootv/photos/shemena-ethiopia/529117743936943/.
[Accessed 11 may 2021].

[34] M. Kassaw, "Competitiveness Model Development for Ethiopian Traditional Fashions in

the Global Market," Addis Ababa University, vol. I, 2013.

[35] Shahriare Mahmood, Pekka Kess, "An overview of demand management through demand
supply chain in fashion industry," INTERNATIONAL JOURNAL OF MANAGEMENT
SCIENCE AND BUSINESS ADMINISTRATION, vol. 2, no. 12, pp. 7- 9, 2014.

[36] N. P. a. R. S., "Artificial Intelligence: A Modern Approach (3)," in Artificial Intelligence,

Prentice Hall, 2010, p. 77.

[37] McCulloch, W. S., & Pitts, W., A logical calculus of the ideas immanent in nervous
activity. The Bulletin of Mathematical Biophysics, 1943, p. 115–133.

83
[38] L. D. a. D. Yu, "Deep Learning: Methods and Applications," in Foundations and Trends
in Signal Processing, vol. 7, oct-2014, p. 428–434.

[39] M. Z. Baker, "Generation of Synthetic Images with Generative Adversarial Networks," p.

22, 2018.

[40] F. Rosenblatt, The Perceptron, a Perceiving and Recognizing Automaton Project Para.,
Cornel Aeronautical Laboratory, jan 1957.

[41] I. Goodfellow, Y. Bengio, and A. Courville, "Deep Learning," MIT Press, 2016.

[42] L. B. Y. B. a. P. H. Y. LeCun, "Gradient-Based Learning Applied to Document

Recognition," in in Proceedings of IEEE, California, USA, , November 1998..

[43] J. S. Sepp Hochreiter, "Long short-term memory," Neural Computation, pp. 1735-1780,
1997.

[44] R. T. M. a. Y. B. Pascanu, "On the difficulty of training Recurrent Neural Networks,"

2012.

[45] D. P. a. M. W. Kingma, "Auto-Encoding Variational Bayes," 2013.

[46] Alexander NILSSON, Martin THÖNNERS, "A Framework for Generative Product
Design Powered by Deep Learning and Artificial Intelligence," June 14, 2018.

[47] Wikipedia, "Deep learning," 12 September 2021. [Online]. Available:

https://en.wikipedia.org/wiki/Deep_learning. [Accessed 16 September 2021].

[48] D. H. a. T. Wiesel, "Receptive Fields, Binocular Interaction and Functional Architecture in

the Cat’s Visual Cortex," Journal of Physiology, vol. 160, pp. 106-154, 1962.

[49] D. J. S. M. a. D. W. Rezende, "Stochastic Backpropagation and Approximate Inference in

Deep Generative Models," 2014.

[50] I. Goodfellow, "NIPS 2016 Tutorial: Generative Adversarial Networks,"

goodfellow2017nips, 2017.

[51] S. Zhou, "Build Basic Generative Adversarial Networks (GANs)," DeepLearning.AI,

[Online]. Available: https://www.coursera.org/learn/build-basic-generative-adversarial-
networks-gans/home/info. [Accessed 14 aug 2021].

[52] Lele Chen,Justin Tian, Guo Li, Cheng-Haw Wu, Erh-Kan King, Kuan-Ting Chen, Shao-
Hang Hsieh, Chenliang Xu, "Tailorgan: Making user-defined fashion designs," 2020 IEEE
Winter Conference on Applications of Computer Vision (WACV), 2020.

[53] Alec Radford, Luke Metz, Soumith Chintala, "Unsupervised Representation Learning with
Deep Convolutional Generative Adversarial Networks," arXiv preprint, 2015.

84
[54] T. P. P. I. A. A. E. Jun-Yan Zhu, "Unpaired Image- to- Image Translation using Cycle
Generative Adversarial Networks," Regular, vol. 9, no. 6, pp. 380-385.

[55] S. L. A. Tero Karras, "A Style-Based Generator Architecture for Generative Adversarial
Networks," 29-mar-2019.

[56] A. L.-C. a. H. T.-M. Carlos Tejeda-Ocampo, "Improving Deep Interactive Evolution with
a Style-Based Generator for Artistic Expression and Creative Exploration," Entropy 2021,
Carlos Tejeda-Ocampo , Armando López-Cuevas and Hugo Terashima-Marin.

[57] Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen, "Progressive Growing of GANs
for Improved Quality, Stability, and Variation," Neural and Evolutionary Computing,
2018.

[58] S. Bagai, "Artificial Intelligence: Automating the Fashion & Textiles Industry," in
Fashion & Textile Industry 4.0 - Opportunities & Challenges for Education 4.0, India,
2021.

[59] C. L. H. a. S. F. Ng, "Predicting seam performance of commercial woven fabrics using

multiple logarithm regression and artificial neural networks," Textile Research Journal 79,
vol. 18, 2009.

[60] Shizhan Zhu, Sanja Fidler, Raquel Urtasun, "Be Your Own Prada: Fashion Synthesis with
Structural Coherence," in Proceedings of the IEEE International Conferenceon Computer
Vision, 2017.

[61] L. Liu, H. Zhang, Y. Ji, and Q. M. Jonathan Wu, "Toward AI fashion design: An
Attribute-GAN model for clothing match,," Neurocomputing,, vol. 341, pp. 156-167,
2019.

[62] Y. L. a. X. Y. a. X. H. a. N. J. a. K. J. a. J. Lu, "A deep learning based interactive

sketching system for fashion images design," 2020.

[63] Y. &. S. P. &. H. K. &. J. S. Jeon, "FashionQ: An AI-Driven Creativity Support Tool for
Facilitating Ideation in Fashion Design," pp. 4-10, 2021.

[64] S. S. C.-Y. C. S. C. H. J. L. Wen-Huang Cheng, "Fashion Meets Computer Vision: A

Survey," ACM Computing Surveys, vol. 54, no. 4, pp. 1-41, 2021.

[65] R. G. A. Chattopadhyay, "Artificial neural networks: applications to textiles. Textile

Progress," p. 1–46, 2004.

[66] S. R. J. M. Mani Naik, "Fashion Advisor using Generative Adversarial Networks,"

International Research Journal of Engineering and Technology (IRJET), vol. 07, no. 05,
pp. 1630-1642, 2020.

[67] Hao Dong, Simiao Yu, Chao Wu, and Yike Guo, "Semantic image synthesis via
adversarial learning Computer Vision.," in IEEE International Conference on, 2017.

85
[68] Al-Halah, Z., Stiefelhagen, R., & Grauman, K., "Fashion forward: Forecasting visual
style in fashion," 2017 IEEE International Conference on Computer Vision (ICCV).

[69] Wenhui Yu, Huidi Zhang, Xiangnan He, Xu Chen, Li Xiong, and Zheng Qin, "Aesthetic-
based clothing recommendation," in Proceedings of the 2018 World Wide Web
Conference on World Wide Web - WWW '18, 2018.

[70] P. F. Masciarelli, "Artificial Intelligence and the Fashion Industry," LUISS Department of
Economics and Finance, 2018/2019.

[71] J. X. S. L. H. X. X. Z. S. Y. Luoqi Liu, "Wow! You Are So Beautiful Today!," ACM

Transactions on Multimedia Computing, Communications, and Applications, vol. 11, no.
1s, pp. 1-22, 2014.

[72] Kuan-Ting Chen, Jiebo Luo, "When fashion meets big data," Proceedings of the 26th
International Conference on World Wide Web Companion - WWW '17 Companion, 2017.

[73] Jashim Uddin Ahmed,Md. Humayun Kabir, asim Uddin, "Sadakalo: Marketing of
Traditional Fashion in the Modern Fashion Industry," June 2014.

[74] C. Joy, "Evolution of Commerce: How AI is disrupting the Fashion industry," UNBXD,
Dec. 2017. [Online]. Available: http://unbxd.com/blog/ai-‐fashion-‐industry. [Accessed 14
September 2021].

[75] Lai, HuiJuan, Lee, SungWon, "The Application of Artificial Intelligence and VR
Technology in Clothing Store Display Design," vol. PP, 2020/08/31.

[76] S. Verma, "Effect of Visual Merchandising Display on Consumer Buying Behavior

towards Intimate Apparel," in Proceedings of online National Conference on Fashion
Apparel and Textile (NCFAT’20), Swami Vivekanand Subharti University Meerut India,
2020.

[77] S. Fernandes, "Influence of Visual Merchandising on the Consumer," in Proceedings of

online National Conference on Fashion Apparel and Textile (NCFAT’20)Fashion &
Textile Industry 4.0 -, India, 2020.

[78] D. Bradley, T. Popa, A. Sheffer, W. Heidrich, and T. Boubekeur, "Markerless garment

capture.," ACM Trans., p. 99:1–99, Aug.2008..

[79] A. G. D. M. K. E. G. a. M. A. F. Berthouzoz, "Parsing sewing patterns into 3d garments.,"

ACM Trans., July 2013. .

[80] C. C. L. W. a. X. J. Y. Meng, "Flexible shape control for automatic resizing of apparel

products.," Comput. Aided Des., p. 68–76, Jan. 2012..

[81] R. A. G. a. I. K. N. Neverova, "Dense pose transfer," in In The European Conference on

Computer Vision (ECCV), September 2018.

86
[82] E. S. a. T. D. J. Long, "Fully convolutional networks for semantic segmentation," in In
2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015.

[83] X. C. Q. F. K. G. a. P. T. B. Zhou, "Garment modeling from a single image," Computer

Graphics Forum, p. 85–91.

[84] E. I. G. P.-M. Verica Lazova, "360-Degree Textures of People in Clothing from a Single
Image," Computer Vision and Pattern Recognition, vol. 1, pp. 6-9, 20 Aug 2019.

[85] Z. W. Z. W. R. Y. a. L. S. D. X. Han, "Viton: An image-based virtual try-on network," in

In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2018.

[86] L. K. WEI GONG, "Aesthetics, Personalization and Recommendation: A survey on Deep

Learning in Fashion," vol. 00, no. 0, pp. 1-38, 2020.

[87] J. L. Y. C. T. S. S. Y. a. J. F. J. Zhao, "Understanding Humans in Crowded Scenes: Deep

Nested Adversarial Learning and A New Benchmark for Multi-Human Parsing," In ACM
M, 2018.

[88] K.-M. H. S.-Y. W. L.-W. T. H.-H. S. a. W.-H. C. Hung-Jen Chen, "BeautyGlow: On-
demand Makeup TRANSFER framework with Reversible Generative Network,"
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019),
2019.

[89] Q. L. C. G. a. Z. S. Y.R. Cui, "Fashiongan: Display your fashion design using conditional
generative adversarial nets.," Computer Graphics Forum 37, vol. 7, pp. 109-119, 2018.

[90] Y. Ge, R. Zhang, X. Wang, X. Tang, and P. Luo, "DeepFashion2: A versatile benchmark
for detection, pose estimation, segmentation, and re-identification of clothing images,"
2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019.

[91] IVONA TAUTKUTE, TOMASZ TRZCINSKI, ALEKSANDER SKORUPA, LUKASZ

BROCKI and KRZYSZTOF MARASEK, "DeepStyle: Multimodal Search Engine for
Fashion and Interior Design," IEEE Access, vol. 7.

[92] Ziwei Li, Ping Luo, Shi Qiu, Xiaogang Wang, Xiaoou Tang, "DeepFashion: Powering
Robust Clothes Recognition and Retrieval with Rich Annotations," 2019 IEEE/CVF
Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019.

[93] Ziad Al-Halah, Rainer Stiefelhagen, Kristen Grauman, "Fashion Forward: Forecasting
Visual Style in Fashion," 2017 IEEE International Conference on Computer Vision
(ICCV), 2017.

[94] Negar Rostamzadeh, Seyedarian Hosseini, Thomas Boquet, Wojciech Stokowiec, Ying
Zhang, Christian Jauvin, Chris Pal, "Fashion-Gen: The Generative Fashion Dataset and
Challenge," 2018.

87
[95] Sun, Hai Wang, Ping, Chang, Yifan, Wang, Hailei, Li, Wenbo, "HRPGAN: A GAN-based
Model to Generate High-resolution Remote Sensing Images," IOP Conference Series:
Earth and Environmental Science, pp. 1755-1315, 2020/01/17.

[96] Ming-Yu Liu and Xun Huang and Jiahui Yu and Ting-Chun Wang and Arun Mallya,
"enerative Adversarial Networks for Image and Video Synthesis: Algorithms and
Applications," Proceedings of the IEEE, 2020.

[97] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros, "Image-to-image
translation with conditional adversarial networks," in IEEE conference on computer vision
and pattern recognition, 2017.

[98] Kristen Vaccaro, Tanvi Agarwalla, Sunaya Shivakumar, and Ranjitha Kumar, "Designing
the Future of Personal Fashion," 2017.

[99] Chaitanya Patel, Zhouyingcheng Liao, Gerard Pons-Moll, "TailorNet: Predicting Clothing
in 3D as a Function of Human Pose, Shape and Garment Style," 2020 IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), 2020.

[100] Natsumi Kato, Hiroyuk Osone, Kotaro Oomori, Yoichi Ochiai, Chun Wei Ooi, "GANs-
based Clothes Design: Pattern Maker Is All You Need to Design Clothing," Proceedings
of the 10th Augmented Human International Conference 2019, 2019.

[101] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, "Image-to-image translation with conditional
adversarial networks," in ” in 2017 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2017.

[102] Phillip Isola and Jun-Yan Zhu and Tinghui Zhou and Alexei A. Efros, "Image-to-Image
Translation with Conditional Adversarial Networks," Computer Vision and Pattern
Recognition, 2018.

[103] Z. YI, "Text-to-image Synthesis for Fashion Design," Master in Information and Network
Engineering, vol. 1, pp. 1-78, 2019.

[104] Zhenliang He and Wangmeng Zuo and Meina Kan and Shiguang Shan and Xilin Chen,
"AttGAN: Facial Attribute Editing by Only Changing What You Want," 2018.

88
Appendices

Appendix A: Scrape Parameters

Search Terms
Fashion
Fashion Design
Fashion Sketches
Clothing
Garments
Ethiopian Fashion
Ethiopian Traditional Clothes
Habesha lebs
Habesha Garment
Habesha Clothing
Shiro meda Clothing
Habesha Fashion Design
Oromo culture
Oromo Cultural clothes
Habesha Cultural Clothes
Ethiopian cultural clothes
Next fashion design
Bahirdar University Fashion Design
Fashion Designs
Habesha dresses
Habesha kemis
Hewan fashion design
Ethiopian Dresses
TABLE A.1: The strings to search for

89
Appendix B: Participant Interview questions

90
91
Appendix C: Original Sample Dataset

92
Appendix D: Sample Results from the Training of StyleGAN

93
Appendix E: Sample Results from the Training of StyleGAN2AdaIn

94
95