A Project report on

ENHANCING RICE QUALITY ASSESSMENT

THROUGH INTEGRATED NEURAL NETWORKS:
A SYNERGISTIC APPROACH WITH CNN AND
CAEs
Submitted in partial fulfillment of the requirements
for the award of the degree of

BACHELOR OF TECHNOLOGY
in

COMPUTER SCIENCE & ENGINEERING

By

M. ANITHA (214G5A0503)

Under the Guidance of

Mr. P. Veera Prakash [Link]., (Ph.D)
Assistant Professor

Department of Computer Science & Engineering

SRINIVASA RAMANUJAN INSTITUTE OF TECHNOLOGY

(AUTONOMOUS)
Rotarypuram Village, B K Samudram Mandal, Ananthapuramu - 515701

2023-2024
SRINIVASA RAMANUJAN INSTITUTE OF TECHNOLOGY
(AUTONOMOUS)
(Affiliated to JNTUA, Accredited by NAAC with ‘A’ Grade, Approved by AICTE, New Delhi &
Accredited by NBA (EEE, ECE & CSE)
Rotarypuram Village, BK Samudram Mandal, Ananthapuramu-515701

DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

Certificate
This is to certify that the Project report entitled ENHANCING RICE QUALITY
ASSESSMENT THROUGH INTEGRATED NEURAL NETWORKS: A
SYNERGISTIC APPROACH WITH CNN AND CAEs is the bonafide work carried
out by M. Anitha bearing Roll Number 214G5A0503 in partial fulfilment of the
requirements for the award of the degree of Bachelor of Technology in Computer
Science & Engineering during the academic year 2023 - 2024.

Project Guide Head of the Department

Mr. P. Veera Prakash [Link]., (Ph.D) Mr. P. Veera Prakash [Link]., (Ph.D)
Assistant Professor Assistant Professor

Date: External Examiner

Place: Rotarypuram
 DECLARATION

I’m Ms. M. Anitha bearing reg no: 214G5A0503 student of SRINIVASA

RAMANUJAN INSTITUTE OF TECHNOLOGY, Rotarypuram, hereby declare
that the dissertation entitled “ENHANCING RICE QUALITY ASSESSMENT
THROUGH INTEGRATED NEURAL NETWORKS: A SYNERGISTIC
APPROACH WITH CNN AND CAEs” embodies the report of my project work
carried out by me during IV year Bachelor of Technology under the guidance of
Mr. P. Veera Prakash, Assistant Professor & Head Of the Department, Department
of CSE and this work has been submitted for the partial fulfillment of the
requirements for the award of the Bachelor of Technology degree.

The results embodied in this project have not been submitted to any other
University of Institute for the award of any Degree or Diploma.

M. ANITHA Reg no: 214G5A0503

 VISION & MISSION OF THE INSTITUTION
Vision:
To become a premier Educational Institution in India offering the best teaching and
learning environment for our students that will enable them to become complete
individuals with professional competency, human touch, ethical values, service
motto, and a strong sense of responsibility towards environment and society at large.

Mission:

 Continually enhance the quality of physical infrastructure and human resources

to evolve in to a center of excellence in engineering education.

 Provide comprehensive learning experiences that are conducive for the students
to acquire professional competences, ethical values, life-long learning abilities
and understanding of the technology, environment and society.

 Strengthen industry institute interactions to enable the students work on realistic

problems and acquire the ability to face the ever changing requirements of the
industry.

 Continually enhance the quality of the relationship between students and faculty
which is a key to the development of an exciting and rewarding learning
environment in the college.

VISION & MISSION OF THE DEPARTMENT OF CSE

Vision:
To evolve as a leading department by offering best comprehensive teaching and
learning practices for students to be self-competent technocrats with professional
ethics and social responsibilities.

Mission:
DM 1: Continuous enhancement of the teaching-learning practices to gain
profound knowledge in theoretical & practical aspects of computer science
applications.
DM 2: Administer training on emerging technologies and motivate the students
to inculcate self-learning abilities, ethical values and social consciousness to
become competent professionals.
DM 3: Perpetual elevation of Industry-Institute interactions to facilitate the
students to work on real-time problems to serve the needs of the society.
 ACKNOWLEDGEMENT
The satisfaction and euphoria that accompany the successful completion of
any task would be incomplete without the mention of people who made it possible,
whose constant guidance and encouragement crowned our efforts with success. It is a
pleasant aspect that I have now the opportunity to express my gratitude for all of
them.
It is with immense pleasure that I would like to express my in debted gratitude
to my Guide Mr. P. Veera Prakash, Assistant Professor, Computer Science &
Engineering, who has guided me a lot and encouraged me in every step of the project
work. I thank him for the stimulating guidance, constant encouragement and
constructive criticism which have made possible to bring out this project work.
I express my deep felt gratitude to Mr. C. Lakshminatha Reddy, Assistant
Professor and Mr. M. Narasimhulu, Assistant Professor, Project Coordinators
for their valuable guidance and unstinting encouragement enabled me to accomplish
my project successfully in time.
I am very much thankful to Mr. P. Veera Prakash, Assistant Professor &
Head of the Department, Computer Science & Engineering, for his kind support
and for providing necessary facilities to carry out the work.
I wish to convey my special thanks to Dr. G. Bala Krishna, Principal of
Srinivasa Ramanujan Institute of Technology for giving the required information
in doing my project work. Not to forget, I thank all other faculty and non- teaching
staff, and my friends who had directly or indirectly helped and supported me in
completing my project in time.
I also express my sincere thanks to the Management for providing excellent
facilities.
Finally, I wish to convey my gratitude to my family who fostered all the
requirements and facilities that I need.

Anitha M

(214G5A0503)
 ABSTRACT

In the realm of food crops, rice stands as a pivotal staple, and ensuring high-
quality rice is vital for consistent standards in the industry. Traditional methods for
assessing rice quality lack precision, emphasizing the necessity for advanced
techniques. The application of neural networks like MobileNet, specifically CNN
(Convolutional Neural Networks), offers a promising solution to enhance accuracy and
automation in classifying rice grain quality.

Additionally, incorporating the feature learning capabilities of Convolutional

Autoencoders (CAEs) further refines the identification of nuanced variations indicative
of rice quality. Our goal is to establish a robust framework that not only surpasses the
limitations of individual techniques but also delivers optimized and unparalleled
outcomes in the task of determining rice quality. This integrated approach aims to
revolutionize the evaluation process, ensuring a more accurate and efficient assessment
of rice quality for the benefit of human health.

Keywords: Rice quality, Neural networks, CNN (Convolutional Neural Networks),

MobileNet, Convolutional Autoencoders (CAEs), Accuracy, Automation, Classifying,
Robust framework, Optimized outcomes.
 CONTENTS Page No.
List of Figures x-xi
List of Tables xii

Abbreviations xiii
Chapter 1 Introduction 1-4
1.1 Problem Statement 1
1.2 Objectives 2
1.3 Machine Learning 2
1.4 Deep Learning 2
1.5 Introduction to CAE 3
1.6 CAE Essential Elements for reconstruction 3-4
Chapter 2 Literature Survey 5-16
Chapter 3 Methodology 17-21
3.1 Loading Rice image Dataset 17
3.2 Preprocessing 17
3.3 Image Reconstruction 17-18
3.4 Feature Extraction 18-19
3.4.1 Size 18
3.4.2 L/B ratio 18
3.4.3 Color 18
3.4.4 Chalkiness 19
3.4.5 Whiteness 19
3.5 Convolutional Neural Network 20-21
3.5.1 Conv2D Layer 20
3.5.2 Filters 21
3.5.3 Kernel Size 21
3.5.4 Activation Function 21
Chapter 4 Planning 22-26
4.1 Existing System 22
4.1.1 Disadvantages of the Existing System 22
4.2 Proposed System 22-23
4.2.1 Advantages of Proposed System 22
4.3 Convolutional Neural Network Model 23
4.3.1 Pooling Layers 23
4.3.2 Convolutional Layers 23
 4.4 Objectives of Planning 24-25
4.4.1 Scope 24
4.4.2 Performance 24
4.4.3 Cost 24
4.4.4 Time 24
4.5 Modules 25
4.6 Functional and Non-Functional Requirements 25
4.6.1 Functional Requirements 25
4.6.2 Non-Functional Requirements 25
4.7 System Requirements 25
4.7.1 Software Requirements 25
4.7.2 Hardware Requirements 25
4.8 Test case 26
Chapter 5 Design 27-30
5.1 Architecture Diagrams 27-28
5.2 Data Flow Diagrams 29
5.2.1 Rules for creating DFD 29
5.2.2 Levels of DFD 29

5.3 User Interface Diagrams 30

Chapter 6 Implementation 31-38
6.1 Experimental Setup 31
6.2 Image down sampling 31
6.3 Technologies and Libraries used 31-33
6.3.1 Python 31
6.3.2 Google colab
32
6.3.3 Pandas
6.3.4 NumPy 32
6.3.5 CV2 32
6.3.6 Keras and TensorFlow 33
6.3.7 Matplotlib 33
6.3.8 PIL
33
6.3.9 SkLearn
33
6.4 Workflow 34-35
6.4.1 Image Loading and Preprocessing 34
6.4.2 CAE Implementation
34
6.4.3 Evaluation Metrics
6.4.4 Visualization 34
6.4.5 Image Quality Evaluation 35
6.4.6 CNN Model training and Evaluation 35
6.4.7 Transfer Learning 35
6.5 Code Snippets 36-39
Chapter 7 Results 40-48
7.1 Comparison between original and 40-41
reconstructed images 40
7.1.1 MSE
40
7.1.2 SSIM
7.1.3 PSNR 41
7.1.4 MAE 41
7.2 Graphs 41
7.2.1 CNN Graphs 41
7.2.2 MobileNet Graphs
42
7.3 Comparison of Feature result 44-48
CONCLUSION AND FUTURE WORK 49
REFERENCES 50-51
PUBLICATION PAPER
PARTICIPATION CERTIFICATE
 LIST OF FIGURES

Fig. No. Description Page No.

1.1 Convolutional Autoencoders Reconstruction 4
4.1 CNN Model 23
4.2 Input Image 26
4.3 Output Image 26
5.1 Architecture of Proposed Method 27
5.2 Use Case Diagram 28
5.3 DFD-0 30
5.4 DFD-1 30
5.5 User Interface diagram 30
6.1 CAE Model 36
6.2 Defining CNN Model 37
6.3 CNN Sequential Main Model 37
6.4 CNN Compilation Model 37
6.5 CNN graph between Training and Validation accuracy 38
6.6 CNN graph between Training and Validation precision 38
6.7 CNN graph between Training and Validation recall 38
6.8 MobileNet Sequential Main Model 38
6.9 MobileNet Compilation Model 39
6.10 MobileNet graph between Training and Validation 39
accuracy
6.11 MobileNet graph between Training and Validation 39
precision
6.12 MobileNet graph between Training and Validation 39
recall
7.1 The graphs (a), (b), (c), and (d) represent the MSE, 40
SSIM, PSNR, and MAE respectively, indicating the
differences between the original and reconstructed
images.
7.2 Comparison of CNN accuracy 41
7.3 Comparison of CNN precision 42
7.4 Comparison of CNN recall 42
7.5 Comparison of MobileNet accuracy 42
7.6 Comparison of MobileNet precision 43
7.7 Comparison of MobileNet recall 43
7.8 Size Feature comparison between original and real 44
time images

x
7.9 Chalkiness Feature comparison between original and 45
real time images
7.10 Whiteness Feature comparison between original and 45
real time images
7.11 Color Feature comparison between original and real 45
time images
7.12 Home page 46
7.13 Contact us page 46
7.14 About us page 47
7.15 Upload image 48
7.16 Classification result 48

xi
 LIST OF TABLES

Table No. Title Page No.

3.1 Features of reconstructed images 19
4.1 Modeling steps in design of each modules 24

xii
 LIST OF ABBREVIATIONS

CAE Convolutional Autoencoders

CNN Convolutional Neural Networks

DL Deep learning
MAE Mean Absolute Error

MSE Mean Square Error

PSNR Peak Signal Noise Ratio
SSIM Structural Similarity Index

KNN K Nearest Neighbour

ReLU Rectified Linear Unit

xiii
Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

CHAPTER 1
INTRODUCTION

In the landscape of food production, the significance of rice as a primary staple

necessitates a keen focus on maintaining and enhancing its quality standards. The
traditional methods employed for evaluating rice quality often fall short in terms of
precision and efficiency, prompting the exploration of more advanced and nuanced
approaches. In response to this, our project introduces a novel methodology that
harnesses the capabilities of Convolutional Autoencoders (CAEs) and Convolutional
Neural Networks (CNN) for the detection and classification of rice quality.

The pivotal role of CNN lies in its adeptness at object identification, specifically
tailored to discern and classify various characteristics associated with rice grains.
Simultaneously, the integration of Convolutional Autoencoders contributes to the
project's depth by facilitating the extraction of intricate features and patterns relevant to
rice quality. This combination of CNN and CAEs creates a robust and comprehensive
framework for the precise identification of subtle variations in rice quality.

Our project seeks not only to address the shortcomings of conventional methods
but also to set a new standard for accuracy and efficiency in rice quality assessment. By
capitalizing on the synergies between CAEs and CNN, MobileNet model we aim to
surpass the limitations of individual techniques, offering a sophisticated and automated
solution for the rice industry. This endeavor aligns with our broader goal of ensuring
the consistency of high-quality rice products, thereby contributing to improved
standards in the food sector and, ultimately, to the well-being of consumers.

1.1 Problem Statement

The Rice Industry faces a significant challenge in accurately and efficiently
detecting quality attributes, such as broken grains, discolorations, and other anomalies
within rice grains.
Traditional methods has lack precision and automation, hindering consistent
quality assessment and decision-making in rice production processes.

Computer Science & Engineering, SRIT Page 1 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Utilizing Convolutional Neural Networks (CNNs) in conjunction with appropriately

curated datasets, we aim to achieve accurate and high-precision classification of rice
grain quality attributes.

1.2 Objectives
To accomplish the project's purpose, the following particular objectives have been
established.
i. To create a Convolutional Autoencoder (CAE)-based solution for rice quality
detection, which aims to capture more complex, abstract features from data and
offers better feature learning capabilities and potential for improved
performance.
ii. To develop an automated rice quality detection system using Convolutional
Neural Networks(CNN) to accurately identify and classify quality attributes
within rice grains.

1.3 Machine Learning

Machine Learning is a field of inquiry to understanding and building methods that
'learn', that is, method that leverage data to improve performance on some set of tasks.
Machine learning algorithms build a model based on sample data, known as training
data, in order to make predictions or decisions without being explicitly programmed to
do so. Machine Learning algorithms are used in a wide variety of applications, such as
in medicine, email filtering, speech recognition, agriculture, computer vision etc.

1.4 Deep Learning

Deep learning is a subset of machine learning, which is essentially a neural network
with three or more layers. These neural networks attempt to simulate the behavior of
the human brain from matching its ability allowing it to learn from large amounts of
data. While neural network with a single layer can still make approximate predictions,
additional hidden layers can help to optimize and refine for accuracy. There are
Convolutional Autoencoders and Convolutional Neural Networks. These two CNNs
and CAEs fall under the umbrella of deep learning, which encompasses a wide range of
neural network architectures designed for various applications, such as image
recognition, natural language processing, and more.

Computer Science & Engineering, SRIT Page 2 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

1.5 Introduction to CAE

Convolutional Autoencoders (CAEs) are a variant of autoencoders specifically
designed for handling structured grid data, such as images. Unlike traditional
autoencoders, which fully connect input and output layers, CAEs leverage
convolutional layers for more effective feature extraction and spatial hierarchies. The
encoder uses convolutional operations to downsample the input data, capturing
hierarchical features. The decoder then employs transposed convolutions to reconstruct
the original input from the learned representation. CAEs are widely employed in
computer vision tasks, including image denoising, generation, and feature learning, due
to their ability to effectively capture spatial relationships and patterns in visual data.

The CAE architecture consists of an encoder network that progressively reduces the
spatial dimensions of the input image, capturing essential features. The decoder network
then reconstructs the original image from this compressed representation. The use of
convolutional layers allows the network to learn spatial hierarchies in a more adaptive
manner compared to fully connected layers. This iterative process enhances the
network's ability to capture and represent essential features within the images.

1.6 CAEs Essential Elements for Reconstruction

Convolutional Autoencoders (CAEs) primarily consist of two main components: the
encoder and the decoder.

1. Encoder: The encoder in Convolutional Autoencoders (CAEs) utilizes convolutional

layers to capture spatial hierarchies and extract features from input images. Employing
filters and pooling layers, it downsamples spatial dimensions, reducing complexity and
generating a compressed representation called the latent space or bottleneck layer.

2. Decoder: In Convolutional Autoencoders (CAEs), the decoder reconstructs the

original input by upsampling the compressed representation. Employing transposed
convolutional layers for upsampling, it increases spatial dimensions. The final decoder
layer produces a reconstructed output, aiming for a close match to the initial input data.

Computer Science & Engineering, SRIT Page 3 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

In Convolutional Autoencoders (CAEs), as shown in Fig. 1.1, both the encoder and
decoder utilize convolutional layers to capture spatial dependencies and recognize
patterns in input data. The convolutional filters within these layers detect features such
as edges and textures, forming a hierarchical arrangement that enables the network to
learn progressively abstract representations from the input images.

Fig. 1.1: Convolutional Autoencoders Reconstruction

Computer Science & Engineering, SRIT Page 4 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

CHAPTER 2
LITERATURE SURVEY

[1] Analysis on Feature Extraction and Classification of Rice Kernels

for Myanmar Rice Using Image Processing Techniques

Authors: Thae Nu Wah, Pann Ei San, Thandar Hlaing

Summary: This research proposes an image processing system for rice grading and
tests its effectiveness using Paw-San rice. KNN classifier is used for classification.
Global Thresholding method is used in this work for image segmentation. Then, a series
of measurements were done using image processing techniques on three classes of Paw-
San rice in Myanmar. The real-field feature of Paw-San rice is percentage of broken
rice contained in the batch. Also the classification of rice can be improved by using for
using more distinct features. The results confirm that the feature extraction and
classification of rice kernel based on image processing for Myanmar Rice. From this
we get accuracy in their range of 83-100%.

Dataset: The dataset used for this study is 3 types of Paw-San Rice, they have taken
data from google.

Methodology: K-NN classifiers used for classification, Flatbed Scan(FBS),

Bwareaopen (one of the morphological operation method),Otsu Method, and Watershed
transformation method are used for classification of rice grain images.

[2] Classification Model of Wheat Grain based on Autoencoder

Author: S. Wentao

Summary: In this study, Deep learning has been used in speech recognition, image
search, image recognition. In the wheat particle database, this model is used by deep
learning (DL) models like autoencoders to categorize and identify photos of damaged
wheat. Deep Learning learning toolbox. It has 2 hidden layers, each of which has 100
nodes. Deep learning highlights the importance of feature learning, Each layer in the
model is transformed by features, and finally gets more abstract features at the top layer,
which solves artificial design problem levy consumption of time and resources. This

Computer Science & Engineering, SRIT Page 5 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

experiment is based on the MATLAB R2015a operating platform, using a wheat

particle recognition database, and using Matlab's. Experiments quantitative statistical
results show that by adjusting the parameters of DL during the training of the
autoencoder network, the training error of samples will be smaller, and the error of test
samples will be reduced, which will improve correct classification and recognition of
diseased wheat images.

Dataset: The data set contains 4 files, which are the images of the training samples, the
labels of the training samples, the images of the test samples, and the labels of the test
samples.

Methodology: FFNN(Feedforward Neural Network) and SAE (Stacked Autoencoder)

Network.

[3] Deep Learning for Rice Quality Classification

Author: N. Hong Son and N. Thai-Nghe

Summary: In this work, two categories of rice (whole rice and broken rice) were
recognized using image processing algorithms and machine learning techniques. This
work proposes an approach for rice quality classification. In this approach, image
processing algorithms and machine learning methods were used to recognize and
classify two difference categories of rice (whole rice and broken rice) based on the rice's
size of the national standard of rice quality evaluation, using Convolutional Neural
Network (CNN). Experimental results for 2000 real images give 93.85% accuracy. The
system also used Support Vector Machines method with HOG features and k-Nearest
Neighbors methods in order to classify and compare the accuracy of those algorithms
which show the results of 85.06% and 84.30% accuracy, respectively. These results
show that rice quality evaluation and classification could be automatically done using
Deep Learning approach.

Dataset: In this study the Datasets are self-collected images about the rice pattern of
Loc Troi 20 breed and captured by 20.7 MP camera of Sony Z1 smartphone.

Methodology: Convolutional Neural Networks(CNN) and Support Vector

Machine(SVM).

Computer Science & Engineering, SRIT Page 6 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Extensions Proposed: As a future challenge ,we can try another edge detection
technique that detects in less time, increasing the mentioned accuracy while decreasing
analysis time.

[4] An automated inspection system for rice seed quality based on deep
learning.
Author: Gao, Y., Wang, W., Zhang, W., & Gao, X.

Summary: This presents an automated inspection system for rice seed quality based on
deep learning techniques. The study aims to address the limitations of traditional rice
seed quality inspection methods, which are time-consuming, labour-intensive, and
subjective. The proposed system uses deep learning algorithms to automatically classify
rice seeds based on their appearance features. The authors collected a dataset of 18,000
rice seed images from different varieties and manually labelled them with their
corresponding quality levels. The study aims to address the limitations of traditional
rice seed quality inspection methods, which are time-consuming, labour-intensive, and
subjective. The proposed system uses deep learning algorithms to automatically classify
rice seeds based on their appearance features.

Dataset: The dataset used for this study is taken from google, kaggle.

Methodology: Convolutional Neural Networks(CNN) for finding the quality of the rice
seeds.

[5] Rice quality analysis using deep learning techniques : A review

Author: Latha, M. N., & Nandhini, R

Summary: In this, discusses the importance of rice quality analysis in the food industry
and the limitations of traditional methods . It then provides an overview of deep learning
techniques, including convolutional neural networks (CNNs), recurrent neural networks
(RNNs), and deep belief networks (DBNs), and their applications in rice quality
analysis. The authors discuss various aspects of rice quality analysis, including size and
shape, texture, and composition, and review the deep learning-based methods that have
been developed for each of these aspects. They also discuss the challenges and future
directions for deep learning-based rice quality analysis, including the need for larger

Computer Science & Engineering, SRIT Page 7 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

and more diverse datasets, more robust and interpretable models, and the integration of
multiple modalities.

Dataset: The dataset used for this study is taken from google, kaggle.

Methodology: It then provides an overview of deep learning techniques, including

convolutional neural networks (CNNs), recurrent neural networks (RNNs), and deep
belief networks (DBNs), and their applications in rice quality analysis.

[6] Machine vision based quality analysis of rice grains

Author: T. G. Devi, P. Neelamegam and S. Sudha

Summary: In this study, this algorithm used simple morphological images processing
methods and it can be used to perform various images based analysis. In machine vision
based testing, we take both physical (grain shape and size) and chemical characteristics
(amylose content, gel consistency) for evaluation and grading of rice grains. Quality
assessment is done by finding 1) the region of boundary and 2) the end points of each
grain by measuring the length, breadth and diagonal size of grain. Majority of the
operations used in this algorithm are based on Matlab commands. In an embedded
imaging based system, images are acquired using CMOS/ CCD based image sensors.
This algorithm is suitable to grade large number of grains efficiently and results were
compared with manual work which provided the percentage of accuracy greater or
much equally to the manual work.

Dataset: The dataset used for this study is taken from google.

Methodology: Get input image -> Binary conversion -> Edge detection -> Finding
dimensions -> Sorting and labelling -> Grading.

Extensions Proposed: As a future challenge , the rice grains are identified with diverse
varieties by extracting the features like length, shape, color and texture properties with
different Machine learning and deep learning methodologies.

Computer Science & Engineering, SRIT Page 8 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

[7] Rice Quality Analysis Using Machine Learning

Author: Annadasu, P. and Jaisharma, K

Summary: In this work, it is applied to create a useful model that facilitates rice grain
grading without requiring more effort. This paved the way for development of
computerized vision in rice quality inspection. In the proposed method both image
processing and machine learning techniques are clubbed to analyze and grade the
quality of rice kernels with the help of Support Vector Machine (SVM) classifier in
python platform. Quality of rice is determined based on the physical and chemical
characteristics like area, length, width, moisture content, whiteness, milling degree etc,.
Segmentation is acquired by many techniques like based segmentation, and
segmentation by clustering. From the results obtained we can conclude that this system
is efficient and cost-effective.

Dataset: In this study the Datasets are self-collected images contains multiple rice
grains in a single image.

Methodology: The model used Support Vector Machine (SVM) for identifying the rice
grains in the taken image.

[8] Comparison of CNN-based deep learning architectures for rice

diseases classification

Author: Md Taimur Ahad, Yan Li, Bo Song, Touhid Bhuiyan

Summary: This study focuses on the performance comparison of convolutional neural

network (CNN) architectures in detecting and localizing rice diseases. The researchers
conducted a rice disease classification comparison of six CNN-based deep-learning
architectures using a database of nine of the most epidemic rice diseases in Bangladesh..
They applied a transfer learning approach to DenseNet121, MobileNetV2, Resnet152V,
Seresnext101, and an ensemble model called DEX to compare the six individual CNN
networks, transfer learning, and ensemble techniques. Transfer learning can increase
the accuracy by 17% from the results obtained by Seresnext101 in detecting and
localizing rice leaf diseases.

Dataset: The dataset used for this study is taken from PlantVillage, google.

Computer Science & Engineering, SRIT Page 9 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Methodology: It has used six CNN-based deep-learning architectures they are

DenseNet121, Inceptionv3, MobileNetV2, resNext101, Resnet152V, and
Seresnext101.

[9] Rice Leaf Disease Recognition using Local Threshold Based

Segmentation and Deep CNN

Author: Anam Islam, Redoun Islam, S. M. Rafizul Haque, S.M. Mohidul Islam,
Mohammad Ashik Iqbal Khan

Summary: The method uses local threshold-based segmentation and the Convolutional
Neural Network (CNN) to segment disease-affected regions of rice leaves. The
proposed method has been applied on three different datasets, including the one created
by the authors, which consists of rice leaf images collected from the Bangladesh Rice
Research Institute (BRRI). Three state-of-the-art CNN architectures, VGG, ResNet, and
DenseNet, have been trained with these three datasets for classifying the diseases. This
study used 4500 images for the experiment. The images were preprocessed by cropping
manually and resized to reduce the time for training. The authors used transfer learning
on AlexNet which is a relatively small scale and old CNN model. They used various
image augmentation techniques on these collected images and managed to get test
accuracy of 91.23%. The classification performance of the proposed method using the
said three CNN architectures for the three datasets has been analyzed and compared.

Dataset: In this study the Datasets are self-collected images, contains three datasets of
Rice Leaf blast, Bacterial Leaf Blight, Sheath Blight.

Methodology: In this, Convolutional Neural Network (CNN) VGG, ResNet, and

DenseNet are used to identify the images of rice leaves which contains the disease.

[10] Convolutional neural network with transfer learning for rice type
classification.

Author: Patel, Vaibhav.

Summary: This paper proposes a deep learning-based method for identifying rice
types, reducing manual labor and error. Two methods are proposed: one uses a deep
convolutional neural network (CNN) trained on segmented rice images, and the other

Computer Science & Engineering, SRIT Page 10 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

uses a combination of a pretrained VGG16 network and the proposed method, using
transfer learning for improved accuracy. We have proposed a transfer learning based
approach for Basmati rice type classification. Though the rice type classification for
Basmati rice group is a harder task than classification of rice types in general or rice
groups, our proposed network with pretrained weights of VGG16 perform better than
the other approaches. The approach can also classify rice grains as broken or fine. The
architecture, pretrained on ImageNet data, significantly improves classification
accuracy despite distinct rice images.

Dataset: The dataset used for this study is taken from google, contains 5 types of rice
Basmati, Arborio, Jasmine, Ipsala, Karacadag.

Methodology: Deep convolutional neural network (CNN) and a pretrained VGG16

network is used for identifying the rice type of the given input images.

[11] CNN and Convolutional Autoencoder (CAE) based real-time

sensor fault detection, localization, and correction

Author: Jana, Debasish & Patil, Jayant & Herkal, Sudheendra & Nagarajaiah, Satish
& Dueñas-Osorio, Leonardo.

Summary: In order to detect, classify, and reconstruct faults in sensor data, the study
introduces a novel deep learning framework for linear systems with time-invariant
parameters. The Convolutional Neural Network (CNN) in the framework is used to
identify the type and existence of faults. For reconstruction, a set of separately trained
Convolutional Autoencoder (CAE) networks is used for each type of fault. The models
demonstrated strong performance for both simulated and experimental datasets with a
single failure. The study empirically proves that the proposed framework performs
better than other state-of-the-art techniques in terms of computational efficiency with
comparable accuracy. Adoption of this framework in online structural health
monitoring applications can lead to minimal disruption to monitoring processes,
reduced downtime for structures and infrastructure, while simultaneously reducing
uncertainty and improving the quality of sensor data for historical records. They
achieved 100% accuracy in faulty sensor localization, over 98.7% accuracy in fault type
classification, and over 98% accuracy in reconstruction.

Computer Science & Engineering, SRIT Page 11 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Dataset: The dataset used for this study is taken from google, contains 50000 images
of 5 types Normal, Missing, Random, Drift, Spiky.

Methodology: The model used Convolutional Autoencoder (CAE) and Convolutional

Neural Network (CNN) to identify the type and existence of faults, this leads to minimal
disruption to monitoring processes, reduced downtime for structures and infrastructure.

[12] Terahertz spectra reconstructed using convolutional denoising

autoencoder for identification of rice grains infested with Sitophilus
oryzae at different growth stages

Author: Pu, Hongbin, et al.

Summary: Convolutional denoising autoencoders, or AEs, have the potential to

perform better when they are created by fusing the CNN model with automatic feature
extraction and automatic parameter optimization. To get rid of contaminants from the
surface and grains without rice, the grains were first soaked in water for thirty minutes.
Additionally, the CNN model in deep learning (DL), a successful feature extraction
technique, has been extensively employed to effectively extract information from THz
spectrum data. In order to further reduce noise and obtain THz spectrum data, the CDAE
along with AE and CNN was built. The AE, an unsupervised data reconstruction
method, mainly consists of an encoder and a decoder, which can better represent the
information of the original data with hidden layers. The CDAE mainly contained an
encoder and a decoder with similar structure, which both have three convolutional
layers with the convolutional kernel of the size 3 × 1 and step size 1. The number of
convolutional kernels of the encoder increased gradually to 32, 64 and 128 respectively,
while the number of convolutional kernels of the decoder gradually decreased to 128,
64 and 32 respectively. Mean square error (MSE) is a loss function that was employed
to assess how well the CDAE performed in recreating the THz spectrum.

Dataset: In this study the Datasets are self-collected images, the dataset contains
multiple rice grains in a image.

Methodology: Convolutional Autoencoders (CAE) and Convolutional Neural Network

(CNN) it is applied to detect Bacterial Spot disease in peach plant, which is caused by

Computer Science & Engineering, SRIT Page 12 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

a bacterium named Xanthomonas Campestris. This model can be used to detect other
plant diseases, as well.

[13] Plant disease detection using hybrid model based on convolutional

autoencoder and convolutional neural network.

Author: Bedi, Punam, and Pushkar Gole.

Summary: Convolutional Neural Networks (CNNs) and Convolutional Autoencoders

(CAEs) are two Deep Learning methods that are widely employed in computer vision
applications because of their superior performance on picture data. These two methods
extract different temporal and spatial elements from image data by using the
convolution procedure. CNNs are utilized for input image classification into their
appropriate classes, whereas CAEs are employed for effectively reducing an image's
dimensionality. This research suggests a novel hybrid model based on CNN and CAE
for automatic plant disease diagnosis. Convolutional Neural Network is a Deep
Learning technique that uses convolution operation instead of simple matrix
multiplication. As compared to other Deep Learning techniques, CNN deals with
images most efficiently. Depicts the architecture of a typical CNN that contains one
Input layer, one Output layer, a set of Convolutional layers, Pooling layers, and Fully
Connected layers. A unique hybrid model is built to automatically detect plant illnesses,
using fewer training parameters than current state-of-the-art systems reported in the
literature. Two deep learning methods are used by this model: CNN and CAE.

Dataset: In this study the Datasets are self-collected images contains different leaf
images of different plants.

Methodology: Convolutional Denoising Autoencoders (CDAE) and Convolutional

Neural Network (CNN) model in deep learning (DL), a successful feature extraction
technique, has been extensively employed to effectively extract information from THz
spectrum data.

Computer Science & Engineering, SRIT Page 13 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

[14] Variety identification of single rice seed using hyperspectral

imaging combined with convolutional neural network.

Author: Qiu, Zhengjun, et al.

Summary: Four types of rice seed were imaged hyperspectrally at two distinct spectral
bands (380–1030 nm and 874–1734 nm). The spectral data were retrieved between 441
and 948 nm (Spectral range 1) and 975 and 1646 nm (Spectral range 2). Different
numbers of training samples were used to build CNN, SVM, and K nearest neighbors
(KNN) models. Models in the Spectral range 2 of KNN, SVM, and CNN outperformed
those in the Spectral range 1 by a little [Link] modified the VGGNet architecture
to handle inputs of one-dimensional spectra. Spectral curve patterns and picture patterns
are similar in a few ways. VGGNet's modular design facilitates easy modification and
extension, and it is chosen because to its outstanding performance in image
classification [Link] model's functionality increased.

Dataset: In this study the Datasets are self-collected images, the rice grain images are
arranged in the form rectangle box and taken as input.

Methodology: Convolutional Neural Network (CNN), Support Vector Machine(SVM),

and K nearest neighbors (KNN) are used to identify the variety of the single rice grain
in the given input image.

[15] Deep-rice: Deep multi-sensor image recognition for grading rice.

Author: Wu, Yiqiang, et al.

Summary: DeepRice is based on a deep learning architecture. To be more precise,

Deep-Rice uses a multi-view CNN architecture to extract discriminative features from
several rice image views and then uses a modified softmax loss function to try and
optimize the CNN parameters. Alongside this deep model, we also constructed a large-
scale rice dataset, known as FIST-Rice, to serve as a foundational resource for food
security research. Every sample is photographed in three distinct lighting scenarios.
Using the FIST-Rice dataset, we compare the suggested Deep-Rice model with the
AdaBoost and SVM techniques. We release the first multi-sensor rice grading dataset
called FIST-Rice(Future Intelligent System Technology), which would greatly advance
the field of rice grading. The outcomes of the experiment suggest that the Deep-Rice

Computer Science & Engineering, SRIT Page 14 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

model performs better under various light intensity settings. Using a modified softmax,
the deep network may learn angularly discriminative features.

Dataset: In this study the Datasets are self-collected images, they are captured using
different cameras and using different angles. Every sample is photographed in three
distinct lighting scenarios

Methodology: Convolutional Neural Network (CNN), and Support Vector

Machine(SVM) used to assess the quality of rice grains.

[16] Automatic rice plant disease recognition and identification using

convolutional neural network.

Author: Rathore, Narendra Pal Singh, and Lalji Prasad

Summary: Computer-aided methods are used to process results from disease detection
quickly and accurately. Convolutional neural networks (CNNs), which automatically
extract characteristics and classify images using fully linked networks, are used by the
deep learning models. We have classified images using CNN architecture. We are
utilizing input photos with the labels "healthy" and "leaf_blast" from the Kaggle dataset.
In this study, 1000 samples of rice crops were used to train CNN using an RGB color
model. We have used CNN model to process the images with the labeled class such as
healthy & leaf blast. Weights are updated automatically in the training process of CNN
that will be able to extract the features of the image. In our proposed architecture CNN
model is having 2 layers. The suggested architectural design for differentiating between
leaf-blast and healthy rice crops. OpenCV, Numpy, and Keras are the deep learning
packages used in the Python development of this model. We employed a batch size of
32, which is a hyper-parameter in deep learning that may be changed. In this paper, we
have trained a Sequential convolutional neural network and obtained a prediction
accuracy of up to 99.61%.

Dataset: In this study the Datasets are collected from kaggle, it contains leaf_blast
images.

Methodology: Convolutional Neural Network (CNN) is used to identify the leaft_blast

images whether they are healthy or infected.

Computer Science & Engineering, SRIT Page 15 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

[17] Enhancing the Classification Accuracy of Rice Varieties by Using

Convolutional Neural Networks.

Author: Tran-Thi-Kim, Nga, et al

Summary: This study uses Convolutional Neural Network (CNN) and Artificial Neural
Network (ANN) models to classify 17 rice grain kinds that are commonly cultivated in
Vietnam. Seventeen rice varieties commonly planted in Vietnam were used in this study
include DT8, HT1, IR4625, IR50404, IR6976, ML48, MO6162, OM108, OM3673,
OM4218, OM429, OM4900, OM5451, OM6976, OM8108, OMCS2012, and RVT.
There were five rice grains in each image. Each variety was represented by 200 images,
from which 100 images were used for training, with the other 100 for testing. The
images for training and testing were fixed for all experiments in this study. Pre-trained
VGG16 and Resnet50 models serve as the foundation for the two CNN models—
modified VGG16 and modified ResNet50. The CNN models were fed the image
dataset, while the ANN was fed the feature dataset. The findings demonstrate that the
classification accuracy of the 17 types of rice was considerably increased by the updated
VGG16 and ResNet50 models. The findings demonstrated that, in comparison to the
SVM, the classification accuracy of a CNN model was much higher. Using the feature
dataset as a model, the ANN produced a 92.82% classification accuracy.

Dataset: In this study the Datasets are self-collected images, the rice grain images were
scanned at a resolution of 2400 dpi. There were five rice grains in each image.

Methodology: Convolutional Neural Network (CNN), Artificial Neural

Network(ANN) is used for rice variety classification.

Computer Science & Engineering, SRIT Page 16 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

CHAPTER 3
METHODOLOGY

Convolutional Autoencoders (CAEs) are a variant of autoencoders specifically

designed for handling structured grid data, such as images. Unlike traditional
autoencoders, which fully connect input and output layers, CAEs leverage
convolutional layers for more effective feature extraction and spatial hierarchies. The
encoder uses convolutional operations to downsample the input data, capturing
hierarchical features. The decoder then employs transposed convolutions to reconstruct
the original input from the learned representation. CAEs are widely employed in
computer vision tasks, including image denoising, generation, and feature learning, due
to their ability to effectively capture spatial relationships and patterns in visual data.

3.1 Loading Rice Images Dataset

Loading the Rice Images Dataset involves retrieving a collection of images

representing various aspects of rice grains. Researchers typically employ data loading
techniques using programming languages like Python and libraries such as TensorFlow
or PyTorch. This dataset is essential for training and evaluating machine learning
models, especially those designed for tasks such as quality assessment or classification
in the context of rice grain analysis.

3.2 Preprocessing

Preprocessing is a crucial step in data preparation for machine learning tasks. It

involves techniques such as normalization, data augmentation, and handling missing
values to enhance the quality and consistency of the dataset. These preprocessing
methods contribute to improved model performance by mitigating noise, ensuring
uniformity, and facilitating better generalization during training.

3.3 Image Reconstruction

In CAEs, image reconstruction involves encoding input images into a latent

representation and then decoding them back into reconstructed images. The objective
is to minimize the reconstruction error, ensuring that the decoded images closely match

Computer Science & Engineering, SRIT Page 17 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

the original input, thus capturing meaningful features in the process.

Encoder: In image reconstruction using CAEs, the encoder is responsible for

transforming input images into a compressed latent representation. This encoded
representation contains essential features extracted from the input, facilitating efficient
storage and subsequent reconstruction during the decoding phase. The encoder's role is
crucial in capturing meaningful information for faithful image representation.

Decoder: In image reconstruction using autoencoders, the decoder reconstructs images

from their compressed latent representations, restoring them to a format similar to the
input. The decoder plays a crucial role in generating faithful reconstructions by
translating the learned features from the latent space back into the original data space.

3.4 Feature Extraction

3.4.1 Size: The size of feature extraction is a crucial factor impacting both model
performance and computational efficiency in machine learning as shown in Table 3.1,
a compact feature representation is essential for reducing memory requirements and
speeding up computations during training and inference. Efficient feature extraction not
only enhances the scalability of models but also facilitates deployment on resource-
constrained devices, contributing to the overall effectiveness of the machine learning
system.

3.4.2 L/B Ratio: The aspect ratio (l/b ratio) in feature extraction plays a pivotal role in
maintaining a balance between capturing long-range and short-range patterns, thereby
influencing the model's capability to learn diverse spatial information as shown in Table
3.1, a higher aspect ratio tends to emphasize long-range dependencies, enabling the
model to grasp broader contextual relationships, while a lower aspect ratio focuses on
finer details and short-range patterns.

3.4.3 Color: The mean color in feature extraction serves as a representative measure of
the average RGB color values within an image, offering valuable insights into color-
based information for machine learning tasks as shown in Table 3.1, analyzing the mean
color enables the model to discern dominant hues and variations, facilitating tasks such

Computer Science & Engineering, SRIT Page 18 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

as image classification and segmentation based on color characteristics. This feature

provides a concise summary of the image's color distribution, enhancing the model's
ability to understand and interpret visual content with a focus on color-related patterns.

3.4.4 Chalkiness: Chalkiness pertains to the degree of lightness or whiteness in an

image, a parameter commonly employed in image analysis to characterize texture or
material properties as shown in Table 3.1, chalkiness measurement allows for the
quantification of textural features related to visual appearance. By incorporating
chalkiness as a feature, machine learning models can effectively discern and classify
variations in material texture, providing valuable insights for tasks such as quality
assessment in crops or food products.

3.4.5 Whiteness: Whiteness signifies the degree of lightness or brightness within an

image, typically quantified using metrics such as luminance or color intensity as shown
in Table 3.1, this parameter holds significance across diverse applications, including
image analysis and quality assessment. Assessing whiteness aids in characterizing the
overall brightness of visual content, providing a valuable feature for machine learning
models to discern and interpret variations in illumination and color intensity in tasks
such as image recognition and quality control.

Table 3.1: Features of Reconstructed images

S. No Size Shape(L/B ratio) Chalkiness Whiteness Mean color (BGR)
1 1790.5 1.942 0.091 24.79 (25.4,24.7,24.7)
2 1685.5 0.666 0.079 24.79 (23.2,22.5,22.6)
3 2025.0 0.796 0.122 29.24 (29.9,29.1,29.1)
4 1870.0 1.169 0.095 25.99 (26.8,25.9,25.8)
5 1872.5 1.037 0.052 23.54 (24.0,23.5,23.4)
6 1579.0 0.767 0.076 22.20 (22.4,22.1,22.1)
7 1509.5 1.875 0.075 21.05 (21.1,21.1,21.1)
8 1703.5 1.340 0.060 21.86 (22.2,21.7,21.8)
9 1943.9 0.057 0.075 25.77 (26.4,25.6,25.7)
10 1760.0 0.730 0.087 24.95 (25.5,24.8,24.9)
11 1748.5 0.646 0.090 24.48 (24.4,24.4,24.4)
12 2082.5 1.850 0.102 33.66 (33.6,33.6,33.6)
13 1855.0 1.131 0.051 24.06 (24.0,24.0,24.0)
14 2027.0 1.055 0.057 25.97 (27.0,26.9,26.9)
15 1783.0 1.842 0.061 24.02 (24.4,23.9,24.2)

Computer Science & Engineering, SRIT Page 19 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Training and Testing datasets were created from the data after the reconstructed images
were gathered and features were retrieved using convolutional Autoencoders.
Convolutional Neural Networks (CNN) and MobileNet are the algorithms we utilized
to train the model. For training and testing sets, arrange image data into the proper
directories. Here, we utilized an unlabeled dataset consisting of 75,000 images divided
into five varieties: Arborio, Basmati, Jasmine, Ipsala and Karacadag. Each rice variety
comprises 15,000 images, and the dataset has been split into three classes - High, Low
and Medium - for each variety. These classes are divided based on the features that were
extracted from the CAEs. For the training and testing datasets, the data is also divided
into three classes.

When loading and preprocessing your image data, you would typically ensure that all
images are resized to these specified dimensions before being fed into the neural
network for training. Image height and Image width, these dimensions are crucial
because they determine the size of the input tensor that the neural network expects. The
values are often chosen based on the characteristics of the dataset and the requirements
of the neural network architecture.

3.5 Convolutional Neural Network

This Convolutional Neural Network architecture involves hyper parameters, plays a

crucial role in shaping the behavior and expressive power of the convolutional layer.
Choosing appropriate values for these parameters depends on the nature of the data, the
complexity of the patterns to be learned, and the overall architecture of the neural
network. Some of the hyper parameters are:

3.5.1 Conv2D Layer: The Conv2D layer serves as a foundational element in

convolutional neural networks (CNNs), specializing in 2D convolutions tailored for
image processing tasks. Operating on input data, this layer utilizes filters to extract
spatial hierarchies and detect patterns within the visual input. By convolving the input
across both dimensions, Conv2D plays a key role in capturing meaningful features,
making it a cornerstone in the architecture of CNNs, especially well-suited for tasks
involving image analysis and recognition.

Computer Science & Engineering, SRIT Page 20 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

3.5.2 Filters: The filters parameter is a critical aspect of Conv2D layers in a neural
network, defining the number of individual filters or kernels that the layer will learn
during training. Each filter specializes in detecting specific patterns or features within
the input data, contributing to the network's ability to capture hierarchical
representations of the input information.

3.5.3 Kernel Size: The kernel size parameter in a Conv2D layer specifies the
dimensions of the convolutional window or filter applied to the input data.A smaller
kernel size captures fine-grained details, while larger sizes help capture broader
patterns, providing a crucial mechanism for controlling the scale and scope of feature
extraction in convolutional neural networks for tasks like image processing and pattern
recognition.

3.5.4 Activation Function: The activation parameter in a Conv2D layer defines the
activation function applied element-wise to the output produced by the convolution
operation. This function introduces non-linearity, enabling the neural network to model
complex relationships and patterns within the input data. Common activation functions
include Rectified Linear Unit (ReLU) for introducing non-linearity and improving
convergence during training. The choice of activation function is crucial for shaping the
model's capacity to capture and represent intricate features in tasks such as image
recognition or object detection.

When building and training a neural network, tuning these hyper parameters often
involves experimentation and iterative refinement to achieve optimal performance on
the specific task at hand. Epochs refer to the number of times the entire training dataset
is processed by the model. Each pass through the entire training dataset is called one
epoch.

Computer Science & Engineering, SRIT Page 21 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

CHAPTER 4
PLANNING

4.1 Existing System:

The existing system performs the classification of rice grains into their respective
divided classes based on the quality taken by the features of its images.

4.1.1 Disadvantages of the Existing System:

1. Results will be based on user input Image.
2. Image preprocessing steps performed were only suitable for Rice grain Images.

4.2 Proposed System:

In this project, we propose a novel approach that combines Convolutional Neural

Networks (CNN) and Autoencoders to address the task of rice quality detection. We
propose an innovative system that integrates Convolutional Neural Networks (CNN),
MobileNet, and Autoencoders to optimize rice quality detection. The proposed system
leverages the fusion of CNN, MobileNet, and Convolutional Autoencoder (CAE)
methodologies. The CAE, implemented using TensorFlow or PyTorch, efficiently
reconstructs rice images for feature extraction. Subsequently, the MobileNet and CNN
models are employed for accurate image classification, capitalizing on MobileNet's
lightweight architecture for efficiency. The novel combination of CAE and CNN,
MobileNet aims to enhance accuracy and efficiency in rice quality assessment, holding
promise for practical applications in the food industry, particularly for quality control
and assurance.
Our application accepts the uploaded image of low resolution and then enhances
and assures the correct detection of Rice grain quality based on reading its features and
model training. It gives output the type of rice grain(high, low, medium) and classifies
if given multiple images at the same time into its classes.
4.2.1 Advantages of Proposed System:
1. Performs detection of real-time images uploaded by users in our application
based on trained data.
2. Classification of Rice grain quality (three classes high, low, and medium). Also,
detect the disease after classification using the CNN algorithm.

Computer Science & Engineering, SRIT Page 22 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

3. Image pre-processing is performed in a generalized manner.

4. The difference between the manual division of classes of rice grain images and
the results given by the Model is so low which indicates the model performance
is accurate.

4.3 Convolutional Neural Network Model(Basic layers)

4.3.1 Pooling Layers: As shown in Fig. 4.1, Pooling layers in a Convolutional Neural
Network (CNN) play a pivotal role in downsampling spatial dimensions while
preserving essential features. Max pooling, a common technique, extracts the maximum
value from a set of neighboring elements, effectively reducing the resolution of feature
maps. This process aids in enhancing computational efficiency, decreasing model
sensitivity to spatial translations, and promoting the extraction of dominant features for
robust pattern recognition in tasks like image classification.

4.3.2 Convolution Layers: As shown in Fig. 4.1, Convolution layers in a

Convolutional Neural Network (CNN) are fundamental components responsible for
extracting features from input data. These layers apply learnable filters to input images,
detecting patterns and spatial hierarchies. By leveraging local connectivity and weight
sharing, convolution layers efficiently capture features, making them well-suited for
tasks like image recognition. The application of multiple filters in parallel enables the
CNN to learn diverse hierarchical representations, facilitating effective feature
extraction and abstraction.

Fig. 4.1: CNN Model

Computer Science & Engineering, SRIT Page 23 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

4.4 Objectives of Planning:

4.4.1 Scope: This project is applicable for the Quality detection of Arborio,
Basmati, Ipsala, and Jasmine Rice images a n d analyzing their Features. It is used
by agricultural practitioners or farmers or anyone user can access our application.

4.4.2 Performance: To acquire an accuracy greater than 98% by using

Convolutional Neural Networks(CNN), by increasing the training data of model
and by giving more images until the line of overfit.

4.4.3 Cost :

 Mode = Organic
 Effort = 16 Person-Month Development
 Time = 4 Months
 Average Staff Required = 4 Persons

4.4.4 Time: On a whole this project takes about 4 months (16 person-months) of
time period for development and testing.

Table 4.1: Modeling steps in design of each modules

As shown in Table 4.1, the timeline outlines modeling steps: requirements

elicitation, modularization has done in month of January, module development,
integration, modeling has done in month of February-March, and final testing,
delivery has done in month of April.

Computer Science & Engineering, SRIT Page 24 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

4.5 Modules:

 Data Pre-processing
 Building CNN model
 Classification, Detection, and Comparison

4.6 Functional and Non-Functional Requirements:

4.6.1 Functional Requirements:

 Scan an image to give input to the system.

 Process and take input of the image, detect the Quality type and compare the
results.
 Display the type quality of the Rice image.
 Show the percentage accuracy of classified images.

4.6.2 Non-Functional Requirements:

 Usability: The application is easy to use with very less complexity of interface.
 Availability: Available within the device.
 Scalability: This model can be trained for other Rice images.
 Performance: Performs well with no sort of failures.

4.7 System Requirements:

4.7.1 Software Requirements:

 Google Colaboratory
 Visual Studio App

4.7.2 Software Requirements:

 Processor: Minimum Intel Core I5

 RAM: Minimum 8 GB
 Hard Disk: Minimum 250 GB
 Operating System: Windows10

Computer Science & Engineering, SRIT Page 25 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

4.8 Test Case:

Input:
The input image, as shown in Fig. 4.2, is a real-time image used for testing purpose.

Fig. 4.2: Input Image

Output:
As shown in Fig. 4.3, the output image describes whether the given rice image is of high
quality, low quality, or medium quality based on the training data. In addition, the output
image includes a probability score, which indicates how accurately the model classified
the quality of the rice image.

Fig. 4.3: Output Image

Computer Science & Engineering, SRIT Page 26 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

CHAPTER 5
DESIGN

5.1 Architecture Diagrams:

An architectural diagram is a visual representation that maps out the physical
implementation for the components of a software system. It shows the general structure
of the software system and the associations, limitations, and boundaries between each
element.

Fig. 5.1: Architecture of Proposed Method

In the proposed system, as shown in Fig. 5.1, the input data undergoes a series of
preprocessing steps to prepare it for analysis. Initially, raw data, often in the form of
images, is fed into the system. Subsequently, image reconstruction and feature
extraction are performed using Convolutional Autoencoders (CAEs). This process

Computer Science & Engineering, SRIT Page 27 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

enables the extraction of meaningful features from the input data, facilitating the
subsequent analysis stages. Following feature extraction, the dataset is split into training
and testing sets to ensure the robustness and generalization of the model.

Once the data is prepared, the training phase begins, wherein deep learning algorithms
such as Convolutional Neural Networks (CNNs) and MobileNets are employed. These
algorithms learn from the extracted features to recognize patterns and make predictions.
During training, the model iteratively adjusts its parameters to minimize the difference
between predicted and actual outputs. After training, the model undergoes testing to
evaluate its performance on unseen data. This evaluation phase involves assessing the
model's accuracy, precision, recall, and other relevant metrics. Finally, the trained
model is ready for deployment, where it can be used for prediction and classification
tasks on new input data.

Fig. 5.2: Use case diagram

Computer Science & Engineering, SRIT Page 28 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

As shown in Fig. 5.2, illustrates a system's workflow. It encompasses data

preprocessing, training utilizing CNN or MobileNet, user authentication, model
evaluation, and result storage. Collaboration between trainers and users is central to the
process.

5.2 Data Flow Diagrams:

The most generic and generally broadest reaching diagram you can make is the
flow diagram. Itis a medium-high level diagram that shows all the pieces of a workflow.
The audience for this type of diagram is generally technical. It may be used to pitch an
idea to an architecture board or describe how a business process works to a developer.
The Data Flow Diagram is a crucial modeling tool (DFD). It models the
system’s components. System processes, data, an external entity interacting with the
system, and information flow within the system are all included in this list. In the Fig.
5.3 and 5.4, the DFD displays how information flows through the system and how it is
transformed through a sequence of transformations. It’s a visual representation of how
data goes from input to output and the transformations that take place along the way.
Specific operations depending on the type of data can be explained by a
flowchart. Data Flow Diagram can be represented in several ways. Data Flow diagrams
are very popular because they help us to visualize the major steps and data involved in
software-system processes.

5.2.1 Rules for creating DFD:

1. The name of the entity should be easy and understandable without any extra
assistance.
2. The processes should be numbered or put in ordered list to be referred easily.
3. The DFD should maintain consistency across all the DFD levels.
4. A single DFD can have maximum processes up to 9 and minimum 3 processes.

5.2.2 Levels of DFD:

DFD uses hierarchy to maintain transparency thus multilevel DFD’s can be created.
Levels of DFD are as follows:

 0-level DFD
 1-level DFD

Computer Science & Engineering, SRIT Page 29 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs


Fig. 5.3: DFD-0

Fig. 5.4: DFD-1

5.3 User Interfaces Diagrams:

Fig. 5.5: User Interface Diagrams

As shown in Fig. 5.5, When a user uploads an image, the system promptly displays the
rice grain quality, categorized as high, medium, or low.

Computer Science & Engineering, SRIT Page 30 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

CHAPTER 6
IMPLEMENTATION

6.1 Experimental Setup:

A 64-bit Windows system was used in the experiment. A high performance cpu
was used for data preprocessing. Utilizing a GPU can significantly accelerate training
times compared to using only a CPU. The CNN and CAE models were built with Keras,
TensorFlow, Numpy, Matplotlib and various other frameworks and packages available
in Python3 through Google Colaboratory environments.

6.2 Image Down sampling:

Images were acquired from Kaggle (a website for datasets). The datasets
contains 5 types of rice(Arborio, Basmati, Ipsala, Jasmine, Karacadag) divided into 5
folders of 15000 images each. These images are given to CAE for image reconstruction.
It consists of convolutional layers for encoding the input image into a latent
representation and then decoding it back to the original image shape. The resolution of
the reconstructed images is [Link] images were used for training and testing a
Convolutional Neural Network model to assess the quality of rice.

6.3 Technologies and Libraries Used:

6.3.1 Python:
Python is used as a scripting language to write the entire codebase, from
importing libraries to defining functions, loops, and conditional statements. Python's
built-in data structures like lists, dictionaries, and tuples are extensively used for storing,
manipulating, and iterating over data. Python's functions and modules are utilized for
organizing code into reusable components. Functions are defined for tasks such as
plotting confusion matrices, computing image quality metrics, and training models.
Modules are used to organize related functions and classes, promoting code readability
and maintainability.

Computer Science & Engineering, SRIT Page 31 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

6.3.2 Google Colab:

Google Colab is a cloud-based platform provided by Google that allows users
to write and execute Python code in a Jupyter Notebook environment, directly in the
browser. It provides free access to GPU and TPU hardware accelerators, which
significantly speed up the training of deep learning models. This is especially useful for
computationally intensive tasks like training convolutional neural networks (CNNs) and
autoencoders. It allows easy access to data stored on Google Drive and provides built-
in commands to upload and download data from the local file system. This makes it
convenient to work with datasets, images, and other files required for the project.
Google Colab comes pre-installed with popular Python libraries and packages,
including TensorFlow, Keras, NumPy, Pandas, Matplotlib, and scikit-learn. This
eliminates the need for manual installation and configuration, enabling users to start
coding immediately. Google Colab seamlessly integrates with other Google services
such as Google Drive, allowing users to store and share notebooks, datasets, and trained
models. This facilitates collaboration among team members and simplifies project
management.

6.3.3 Pandas:
‘Pandas’ is a popular data manipulation library in Python that provides data
structures such as Data Frame and Series, making it easy to manipulate, analyze, and
visualize data. It provides powerful tools for data cleaning, aggregation, filtering, and
transformation, making it a go-to library for data analysis and data wrangling tasks.

6.3.4 Numpy:
NumPy is a library for the Python programming language, adding support for
large, multi-dimensional arrays and matrices, along with a collection of mathematical
functions. It is extensively used in the code for array manipulation, mathematical
computations, and data handling.

6.3.5 CV2:
OpenCV is a library of programming functions mainly aimed at real-time
computer vision. It is used in the code for image processing tasks such as loading images
and manipulating them.

Computer Science & Engineering, SRIT Page 32 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

6.3.6 Keras and TensorFlow:

TensorFlow is an open-source deep learning framework developed by Google.
Keras is an open-source neural network library written in Python. It acts as an interface
for TensorFlow, making it easier to define and train deep learning model. These are
used to define and build deep learning models. In the code, both Convolutional Neural
Networks (CNNs) and Convolutional Autoencoders (CAEs) are constructed using
TensorFlow's Keras API. The Sequential model and its various layers (Conv2D,
MaxPool2D, Dense, GlobalAveragePooling2D, BatchNormalization, Dropout) are
utilized to design the architectures of these models. TensorFlow provides optimization
algorithms and training routines for training deep learning models. In the code, the
compile() method configures the model for training by specifying the optimizer, loss
function, and evaluation metrics. Additionally, the fit() method is used to train the
models using training data generated by ImageDataGenerator.

6.3.7 Matplotlib:
Matplotlib is a plotting library for the Python programming language. It is used
to visualize images, model training/validation metrics, and confusion matrices.
Matplotlib is a widely used plotting library in Python that provides a flexible and
comprehensive API for creating visualizations. It supports a wide range of plots,
including line plots, scatter plots, bar plots, histograms, and more, making it suitable
for data visualization and exploration tasks.

6.3.8 PIL (Python Imaging Library):

PIL is a library for opening, manipulating, and saving many different image file
formats. In the code, PIL is used for loading and processing images before training the
models.

6.3.9 Sklearn(scikit-learn):
Scikit-learn is a machine learning library for Python. It is used for computing
metrics like confusion matrices and classification reports for evaluating model
performance.

Computer Science & Engineering, SRIT Page 33 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

6.4 Work Flow:

6.4.1 Image Loading and Preprocessing:
The code imports necessary libraries such as PIL (Python Imaging Library) for
image manipulation then images are loaded using the [Link]() function from PIL.
After that image dimensions are retrieved to understand the size and channels of the
images. A list of image filenames is generated from the folder path. Images are loaded,
and basic information about their dimensions is printed. Images are displayed using
Matplotlib.

6.4.2 CAE Implementation:

The libraries such as TensorFlow and Keras are imported for deep learning
implementation. A CAE model is defined using the Keras Functional API. The model
consists of convolutional layers for encoding and decoding. The model is compiled with
an optimizer ('adam') and a loss function ('MSE').Image filenames are iterated through.
Images are loaded and resized to a standard size of 128x128 pixels. Images are
normalized (scaled to the range [0, 1]). The CAE model is trained on the input image
for 50 epochs with a batch size of 1. The trained model is used to reconstruct the input
image. Reconstructed images are saved in an output directory. Original and
reconstructed images are displayed side by side using Matplotlib.

6.4.3 Evaluation Metrics:

Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR) are
calculated to evaluate the quality of reconstructed images. MSE is computed using
NumPy to quantify the difference between the original and reconstructed images. PSNR
is computed using the Peak Signal Noise Ratio function from the [Link]
module.

6.4.4 Visualization:
MSE and PSNR values are displayed alongside the original and reconstructed
images for visual comparison. This implementation theory outlines the steps involved
in loading, preprocessing, training, and evaluating the CAE model for image
reconstruction. It also includes visualizations to assess the quality of the reconstructed
images.

Computer Science & Engineering, SRIT Page 34 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

6.4.5 Image Quality Evaluation:

Structural Similarity Index (SSIM) and Mean Absolute Error (MAE) are
computed to further evaluate the quality of reconstructed images. SSIM is calculated
using the structural similarity function from the [Link] module. MAE is
computed using NumPy to measure the average absolute difference between the
original and reconstructed images. SSIM and MAE values are displayed alongside the
original and reconstructed images for visual comparison.

6.4.6 CNN Model Training and Evaluation:

Utilizing TensorFlow and Keras, a Convolutional Neural Network (CNN)
model is constructed for rice quality classification. The CNN model architecture
consists of convolutional layers followed by max-pooling layers and fully connected
layers. The model is compiled with the Adam optimizer and categorical cross-entropy
loss function. Training and validation accuracy, loss, precision, and recall metrics are
monitored and plotted over epochs.

6.4.7 Transfer Learning:

A pre-trained MobileNet model is employed for transfer learning to enhance the
classification model. The MobileNet model is fine-tuned by adding additional layers
and retraining on the rice quality dataset. Similar training and evaluation steps are
performed, and model performance is analyzed through metrics.

Computer Science & Engineering, SRIT Page 35 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

6.5 Code Snippets:

Fig. 6.1: CAE Model

As shown in Fig. 6.1, the code lines represents the CAE model which contains encoder
and decoder with their layers respectively with parameters .Here it is used to reconstruct
the images for further feature extraction.

Computer Science & Engineering, SRIT Page 36 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Fig. 6.2: Defining CNN Model

As shown in Fig. 6.2, the code lines represents pre train model which comes before the
main model . Using pre trained model helps main model to grasp the features effectively
and trained well.

Fig. 6.3: CNN Sequential Main Model

As shown in Fig. 6.3, the code lines represents the Main CNN model (Sequential). Layers
are included with their respective parameters and model is created.

Fig. 6.4: CNN Compilation Model

As shown in Fig. 6.4, the code lines represents the CAE model compilation.

Computer Science & Engineering, SRIT Page 37 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Fig. 6.5: CNN graph between Training and Validation accuracy

Fig. 6.6: CNN graph between Training and Validation Precision

Fig. 6.7: CNN graph between Training and Validation recall

As shown in Fig. 6.5, 6.6, and 6.7, the code lines represents the Classification metrics i.e.,
(Accuracy , Precision and Recall). These are graph plottings between Training images
and validation images. This is for CNN Model Compilation model.

Fig. 6.8: MobileNet Sequential Main Model

Computer Science & Engineering, SRIT Page 38 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

As shown in Fig. 6.8, the code lines represents the MobileNet model which contains their
layers respectively with parameters . It is another main model which is especially used
for mobile embedded applications and for classification tasks.

Fig. 6.9: MobileNet Compilation Model

Fig. 6.10: MobileNet graph between Training and Validation accuracy

Fig. 6.11: MobileNet graph between Training and Validation Precision

Fig. 6.12: MobileNet graph between Training and Validation recall

As shown in Fig. 6.10, 6.11, and 6.12, the code lines represents the Classification metrics
i.e., (Accuracy , Precision and Recall). These are graph plottings between Training images
and validation images for this MobileNet Model.

Computer Science & Engineering, SRIT Page 39 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

CHAPTER 7
RESULTS

7.1 Comparison between Original and Reconstructed Images:

(a) (b)

(c) (d)

Fig. 7.1: The graphs (a), (b), (c), and (d) represent the MSE, SSIM, PSNR, and MAE respectively,
indicating the differences between the original and reconstructed images.

7.1.1 MSE:
As shown in Fig. 7.1 (a), a popular metric for evaluating the average squared
difference between comparable elements of two sets, such as projected and actual
values, is mean squared error, or MSE. It is computed by averaging the squared
differences between predicted and true values. MSE offers a measure of the average
magnitude of errors, where higher values indicate greater overall error in prediction or
reconstruction tasks.

7.1.2 SSIM:
As shown in Fig. 7.1 (b), a metric for comparing two images similarity is called the
Structural Similarity Index (SSIM).It takes into account luminance, contrast, and
structure, providing a more comprehensive measure of perceptual image quality than
some traditional metrics. A higher SSIM value indicates greater similarity between the
original and distorted images, with 1 representing perfect similarity.

Computer Science & Engineering, SRIT Page 40 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

7.1.3 PSNR:
As shown in Fig. 7.1 (c), a typical metric used in image and video compression is
the Peak Signal-to-Noise Ratio (PSNR), which measures the quality of reconstructed
signals. It will be calculated in Decibels. It calculates the ratio of the mean squared error
of the original and reconstructed signals to the highest feasible signal value. Greater
faithfulness and less distortion in the reconstruction are indicated by a higher PSNR.

7.1.4 MAE:
As shown in Fig. 7.1 (d), one statistic used to quantify the average size of deviations
between expected and actual values is mean absolute error (MAE). By averaging the
absolute disparities between the true and anticipated values, it is calculated. MAE offers
a straightforward assessment of the average error magnitude, rendering it especially
valuable in regression and prediction endeavours.

These graphs are used to calculate the loss between original and reconstructed images
after performing the reconstruction using Convolutional Autoencoders.

7.2 Graphs:
7.2.1 CNN Graphs:

Fig. 7.2: Comparison of CNN accuracy

Computer Science & Engineering, SRIT Page 41 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Fig. 7.3: Comparison of CNN precision

Fig. 7.4: Comparison of CNN recall

7.2.2 MobileNet Graphs:

Fig. 7.5: Comparison of MobileNet accuracy

Computer Science & Engineering, SRIT Page 42 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Fig. 7.6: Comparison of MobileNet precision

Fig. 7.7: Comparison of MobileNet recall

These graphs shows that how well the model is performing with CNN and MobileNet.
It shows how well the model is trained on the given images as input and how well the
model is performing on the given images.

In a neural network model, the optimizer is a mathematical algorithm that adjusts the
weights and biases during training to minimize the chosen loss function, guiding the
model towards better performance. The specified metrics during training provide
performance measures for evaluating the model's classification abilities: Accuracy,
Precision, Recall.

Accuracy: It is a vital metric in classification models, measures the ratio of correctly

classified instances to total predictions during both training and testing. Calculated by
dividing correct predictions by total predictions, higher accuracy values. As shown in
Fig. 7.2 and 7.5, for training and testing data, denote superior model performance. These
figures offer insights into the model's effectiveness in classification tasks.

Computer Science & Engineering, SRIT Page 43 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Precision: It evaluates the model's performance on testing and training data. It

represents the ratio of true positives to the sum of true positives and false positives,
indicating the model's reliability in avoiding false positives. As shown in Fig. 7.3 and
7.6, depict precision values, highlighting the model's ability to accurately classify
positives in both datasets and offering valuable insights into its overall performance.

Recall: This metric is crucial in evaluating its performance on both training and testing
data. High recall values indicate the model's effectiveness in identifying actual positive
instances and minimizing false negatives. As shown in Fig. 7.4 and 7.7, depicting recall
values across datasets, offer valuable insights into the model's ability to accurately
identify relevant instances, enhancing its overall performance.

7.3 Comparison of Feature result:

Here the graphs shows the difference between the features extracted from original and
real time images and based on the features, the images are categorized into three classes
High, Low, Medium.

Fig. 7.8: Size Feature comparison between original and real time images

As shown in Fig. 7.8, this graph shows the size feature comparison. As it can be seen that
the difference between orange bar and blue bar is low in most of the cases , hence we can
say the model performance is good and giving accurate results. Here Blue Bar indicates
the size features of original images and Orange bar indicates size features of Real time
images .

Computer Science & Engineering, SRIT Page 44 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Fig. 7.9: Chalkiness Feature comparison between original and real time images

Fig. 7.10: Whiteness Feature comparison between original and real time images

Fig. 7.11: Mean color Feature comparison between original and real time images

The graphs as shown in Fig. 7.8, 7.9, 7.10, and 7.11 illustrate the comparison of features.
These graphs show that even when real-time images are provided to the model by the
user, it can still produce and categorize the images based on how closely the features of
the original and real-time images differ.

Computer Science & Engineering, SRIT Page 45 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Fig. 7.12: Home page

The main page layout of the interface for our project is displayed as shown in Fig. 7.12,
as seen in the images below, three primary modules have been built, each with
corresponding buttons.

Fig. 7.13: Contact us Page

The contact page layout of our project's interface is displayed as shown in Fig. 7.13, as
can be seen, this is the page where users register and establish an account using their
email address and password. On the other side, this is the login page, where users who
already have an account can log in using their credentials.

Computer Science & Engineering, SRIT Page 46 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Fig. 7.14: About us Page

As shown in Fig. 7.14, the user is provided with an overview of our application and
page. This includes a description and an illustration of the project's scope.

Computer Science & Engineering, SRIT Page 47 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

Fig. 7.15: Upload rice image

As shown in Fig. 7.15, user can upload their respective rice image and click the upload
image button.

Fig. 7.16: Classification result

As shown in Fig. 7.16, where the rice image is classified into one of three classes. High,
Low, or Medium. In order to show how accurately the image has been anticipated and
identified, a probability score is also provided.

Computer Science & Engineering, SRIT Page 48 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

CONCLUSION AND FUTURE WORK

In culmination, our project's fusion of Sequential Convolutional Neural

Networks (CNN) and Convolutional Autoencoders (CAEs) signifies a breakthrough in
precision for rice grain quality assessment. By harmonizing CNN's adept object
identification with CAEs' feature learning, the resulting framework exceeds the
limitations of conventional methods. The model's fine-tuning, coupled with data
augmentation, ensures robust adaptability and generalization, validated by impressive
evaluation metrics.

As we deploy this cutting-edge technology into practical applications, it not only

addresses existing challenges but also pioneers a new era of automated and elevated
rice quality standards, crucial for global food security. This success is a testament to the
collaborative efforts of neural network researchers, showcasing the impactful role of
machine learning in revolutionizing the production and quality control of vital food
crops like rice, ensuring both agricultural sustainability and consumer well-being.

Future enhancements to our integrated neural network approach for rice quality
assessment could involve the incorporation of advanced sensor technologies and real-
time data collection during the rice cultivation and processing stages. Implementing a
more dynamic system that adapts to changing environmental conditions and crop
variations could further improve the model's accuracy and robustness. Additionally,
exploring the integration of explainable AI techniques can enhance the interpretability
of the model's decisions, providing valuable insights for stakeholders. Continuous
updates to the dataset and model architecture will ensure adaptability to emerging rice
varieties and quality parameters.

Computer Science & Engineering, SRIT Page 49 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

REFERENCES
[1]. Thae Nu Wah, Pann Ei San, Thandar Hlaing (2018); “Analysis on Feature
Extraction and Classification of Rice Kernels for Myanmar Rice Using Image
Processing Techniques”; International Journal of Scientific and Research Publications
(IJSRP) 8(8) (ISSN: 2250-3153).

[2]. S. Wentao, “Classification Model of Wheat Grain based on Autoencoder” 2020

IEEE International Conference on Artificial Intelligence and Computer Applications
(ICAICA), Dalian, China, 2020, pp. 830-832.

[3]. N. Hong Son and N. Thai-Nghe, “Deep Learning for Rice Quality Classification”,
2019 International Conference on Advanced Computing and Applications (ACOMP),
Nha Trang, Vietnam, 2019, pp. 92-96.

[4]. Gao, Y., Wang, W., Zhang, W., & Gao, X. (2020). “An automated inspection system
for rice seed quality based on deep learning”, Journal of Food Process Engineering,
43(6), e13396.
[5]. Latha, M. N., & Nandhini, R. (2020). “Rice quality analysis using deep learning
techniques: A review”. Computers and Electronics in Agriculture, 174, 105404.
[6]. T. G. Devi, P. Neelamegam and S. Sudha, “Machine vision based quality analysis
of rice grains”, 2017 IEEE International Conference on Power, Control, Signals and
Instrumentation Engineering (ICPCSI), Chennai, India, 2017, pp. 1052-1055.
[7]. Annadasu, P. and Jaisharma, K., 2020. “Rice Quality Analysis Using Machine
Learning”, (No. 4761). EasyChair.
[8]. Md Taimur Ahad, Yan Li, Bo Song, Touhid Bhuiyan, “Comparison of CNN-based
deep learning architectures for rice diseases classification”,Artificial Intelligence in
Agriculture,Volume 9,2023, ISSN 2589-7217.
[9]. Anam Islam, Redoun Islam, S. M. Rafizul Haque, S.M. Mohidul Islam, Mohammad
Ashik Iqbal Khan, “Rice Leaf Disease Recognition using Local Threshold Based
Segmentation and Deep CNN”, International Journal of Intelligent Systems and
Applications(IJISA), Vol.13, No.5, pp.35-45, 2021.
[10]. Patel, Vaibhav. (2017). “Convolutional neural network with transfer learning for
rice type classification”.

Computer Science & Engineering, SRIT Page 50 of 51

Enhancing Rice Quality Assessment through Integrated Neural Networks: A Synergistic Approach with
CNN and CAEs

[11]. Jana, Debasish & Patil, Jayant & Herkal, Sudheendra & Nagarajaiah, Satish &
Dueñas-Osorio, Leonardo. (2022). “CNN and Convolutional Autoencoder (CAE) based
real-time sensor fault detection, localization, and correction”. Mechanical Systems and
Signal Processing. 169. 108723.
[12]. Pu, Hongbin, et al. “Terahertz spectra reconstructed using convolutional denoising
autoencoder for identification of rice grains infested with Sitophilus oryzae at different
growth stages”. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy
(2024): 124015.
[13]. Bedi, Punam, and Pushkar Gole. “Plant disease detection using hybrid model
based on convolutional autoencoder and convolutional neural network”. Artificial
Intelligence in Agriculture 5 (2021): 90-101.
[14]. Qiu, Zhengjun, et al. “Variety identification of single rice seed using hyperspectral
imaging combined with convolutional neural network”. Applied Sciences 8.2 (2018):
212.
[15]. Wu, Yiqiang, et al. “Deep-rice: Deep multi-sensor image recognition for grading
rice”. 2018 IEEE International Conference on Information and Automation (ICIA).
IEEE, 2018.
[16]. Rathore, Narendra Pal Singh, and Lalji Prasad. “Automatic rice plant disease
recognition and identification using convolutional neural network”. Journal of critical
reviews 7.15 (2020): 6076-6086.
[17]. Tran-Thi-Kim, Nga, et al. “Enhancing the Classification Accuracy of Rice
Varieties by Using Convolutional Neural Networks”. International Journal of Electrical
and Electronic Engineering & Telecommunications 12.2 (2023): 150-160.

Computer Science & Engineering, SRIT Page 51 of 51