Convolutional Neural Networks for Image

Classification
Author Details:

Jasmin Praful Bharadiya, Doctor of Philosophy Information Technology,

University of the Cumberlands, USA, [email protected]

Abstract

Deep learning has recently been applied to scene labelling, object tracking, pose estimation, text
detection and recognition, visual saliency detection, and image categorization. Deep learning
typically uses models like Auto Encoder, Sparse Coding, Restricted Boltzmann Machine, Deep
Belief Networks, and Convolutional Neural Networks. Convolutional neural networks have
exhibited good performance in picture categorization when compared to other types of models. A
straightforward Convolutional neural network for image categorization was built in this paper.
The image classification was finished by this straightforward Convolutional neural network. On
the foundation of the Convolutional neural network, we also examined several learning rate
setting techniques and different optimisation algorithms for determining the ideal parameters that
have the greatest influence on image categorization.

Keywords: Convolutional neural network, Deep Learning, Transfer Learning, ImageNet, Image
classification; learning rate, parametric solution

Introduction

Image classification in computer vision is important for our education, jobs, and daily life.
Images are classified using a procedure that includes image preprocessing, image segmentation,
key feature extraction, and matching identification. With the aid of the most modern image
classification techniques, we are now able to acquire image data more quickly than ever before
and put it to use in a number of fields, including face recognition, traffic identification, security,
and medical equipment. In order to address the shortcomings of the conventional approach of
feature selection, feature extraction and classifier have been merged into a learning framework
with the emergence of deep learning. The goal of deep learning is to identify several layers of
representation with the expectation that high-level characteristics will capture the data's more
ethereal semantics. Using Convolutional architectures in image classification is a crucial
component of deep learning. The anatomy of the mammalian visual system serves as inspiration
for convolutional neural network. Hubel and Wiesel suggested a visual structure model based on
the cat visual brain in 1962. For the first time, the idea of a receptive field has been put out. In
1980, Fukushima presented the first hierarchical framework Neocognition would utilise to
analyse pictures. In order to achieve network translation invariance, Neocognition utilised the
local connection between neurons.

There are several deep learning architectures available. Convolutional neural networks, the most
effective and practical deep neural network for this sort of data, were utilised to create the model
reported in this research, a classifier system. As a result, CNNs that have been trained on huge
datasets of pictures for recognition tasks may be used to their advantage by applying these
learning representations to tasks that need less training data.

Since 2006, a variety of techniques have been created to get around the challenges involved in
training deep neural networks. Krizhevsky suggests a traditional CNN architecture Alexnet and
demonstrates a considerable advancement over earlier approaches to the picture classification
job. Numerous initiatives to boost Alexnet's performance have been recommended in light of its
success. VGGNet, GoogleNet, and ZFNet are suggested.

Hierarchical Feature Extraction: CNNs excel at learning hierarchical representations of

images. They consist of multiple layers, including convolutional layers and pooling layers, that
progressively extract features at different levels of abstraction. This hierarchical approach allows
CNNs to capture intricate patterns and structures in images, leading to more accurate
classification.

Key reasons for the significance of CNN

Translation Invariance:

CNNs are designed to be translation invariant, meaning they can recognize patterns regardless of
their location in an image. This is achieved through the use of convolutional layers that apply
filters to an image, detecting features regardless of their position. This property enables CNNs to
classify images regardless of their orientation or position, making them more robust and accurate
in real-world scenarios.

Data Efficiency:

CNNs require fewer training examples than traditional machine learning algorithms. They can
learn from a small number of examples due to their ability to capture relevant features and
generalize to unseen data. This property makes CNNs ideal for scenarios where large amounts of
labeled data are not available.

Transfer Learning:

CNNs are capable of transfer learning, meaning they can learn from one task and transfer that
knowledge to another related task. This is achieved through the use of pre-trained models, which
are trained on large datasets, and can be fine-tuned for specific image classification tasks.
Transfer learning reduces the amount of training data required and can lead to significant
improvements in classification performance.

Scalability:

CNNs are scalable, meaning they can be used for image classification tasks with varying levels
of complexity. This scalability is due to their ability to add or remove layers, adjust the number
of filters in each layer, and change the size of the filters used in convolutional layers. This
flexibility makes CNNs suitable for a wide range of applications, from simple image
classification to more complex tasks such as object detection and segmentation.

To achieve this goal, the objectives below have been specified:

 1. Design a flexible gene encoding scheme of the architecture, which does not constrain the
maximal length of the building blocks in CNNs. With this gene encoding scheme, the
evolved architecture is expected to benefit CNNs to achieve good performance in solving
different tasks at hand.
2. Investigate the connection weight encoding strategy, which is capable of representing
tremendous numbers of the connection weights in an efficient way. With this encoding
approach, the weight connection initialization problem in CNNs is expected to be
effectively optimized by the proposed GA.
3. Develop associated selection (including the environmental selection), crossover, and
mutation operators that can cope with the designed gene encoding strategies of both
architectures and connection weights.
4. Propose an effective fitness measure of the individuals representing different CNNs,
which does not require intensive computational resources.
5. Investigate whether the new approach significantly outperform the existing methods in
both classification accuracy and number of weights.

Methodology of Evaluation

Our research's major goal is to comprehend how effectively networks operate with both static
and real-time video streams. Transfer learning on networks using picture datasets is the initial
stage in the next process. The next stage is to execute transfer learning on networks with picture
datasets. This is followed by testing the next phase. The prediction rate of the same item on still
photos and live video streams is then examined.

The various accuracy rates are noticed, recorded, and shown in the tables provided in subsequent
sections. The third crucial factor for judging the performance was to see if there were any
differences in prediction accuracy between the CNNs used in the study. Videos are utilised as
testing datasets, not as a training dataset, it must be highlighted. As a result, we are searching for
the best picture classifier where the object is the primary attribute for scene category
categorization.
Different layers of the convolutional neural network used are:

Input Layer: The first layer of each CNN used is ‘input layer’ which takes images, resize them
for passing onto further layers for feature extraction.

Convolution Layer: The next few layers are ‘Convolution layers’ which act as filters for
images, hence finding out features from images and also used for calculating the match feature
points during testing.

Pooling Layer: The extracted feature sets are then passed to ‘pooling layer’. This layer takes
large images and shrink them down while preserving the most important information in them. It
keeps the maximum value from each window, it preserves the best fits of each feature within the
window.

Rectified Linear Unit Layer: The next ‘Rectified Linear Unit’ or ReLU layer swaps every
negative number of the pooling layer with 0. This helps the CNN stay mathematically stable by
keeping learned values from getting stuck near 0 or blowing up toward infinity.

Fully Connected Layer: The final layer is the fully connected layers which takes the high-level
filtered images and translate them into categories with labels.

Basic CNN components

Convolutional layer, pooling layer, and fully-connected layer are the three major types of
convolutional neural network layers.
The steps of proposed method are as follows:

1. Creating training and testing dataset: The super classes images used for training is
resized [224,244] pixels for AlexNet and [227,227] pixels GoogLeNet and ResNet50, and
the dataset is divided into two categories i.e. training and validation data sets.
2. Modifying CNNs network: Replace the last three layers of the network with fully
connected layer, a softmax layer, and a classification output layer. Set the final fully
connected layer to have the same size as the number of classes in the training data set.
Increase the learning rate factors of the fully connected layer to train network faster.
3. Train the network: Set the training options, including learning rate, mini-batch size, and
validation data according to GPU specification of the system. Train the network using the
training data.
4. Test the accuracy of the network: Classify the validation images using the fine-tuned
network, and calculate the classification accuracy. Similarly testing the fine tune network
on real time video feeds for accurate results.

Models

There are several intelligent pre-trained CNN; these CNN can transmit learning. Therefore, at its
input layer, it just needs the training and testing datasets. The core layers and methods employed
in the networks' architecture vary. The Inception Modules in GoogleNet execute convolutions of
varying sizes and combine the filters for the following layer. AlexNet, on the other hand, utilises
the output of the preceding layer as its input rather than filter concatenation. Both networks have
undergone independent testing and make use of the Caffe Deep Learning framework's
implementation.
However, as we go further away, neural network training gets challenging and accuracy begins
to saturate before declining. Residual Learning makes an effort to address both of these issues. A
deep convolutional neural network often has many layers that are layered and trained for the
given purpose. At the conclusion of its layers, the network learns a number of low-, mid-, and
high-level characteristics. In residual learning, the network tries to learn some residual rather
than certain characteristics. Residual is just the feature learnt from the layer's input that is
subtracted. ResNet does this through a shortcut connection that connects some (n+x) of the
layer's input straight to another layer. The comparison is made among three existing neural
networks i.e. the AlexNets, Google Nets and ResNet50. The training of existing networks and
the creation of new networks for additional comparison are then followed by the transfer learning
ideas. The new models have the same number of layers as the original models, but their
performance differs greatly from that of the old networks. The tables in the next section provide
the varied accuracy rates that were calculated on the identical photos.

Convolutional Neural Networks (CNNs) for image classification of advancements:

Attention Mechanisms: Recent advancements in CNN architectures have introduced attention

mechanisms, which enable the network to focus on specific regions or features in an image that
are most relevant for classification. You can explore different attention mechanisms, such as
self-attention or spatial attention, and their impact on improving the accuracy and interpretability
of CNN models.

Transformer-Based Architectures: The success of the Transformer model in natural language

processing has led to its adaptation for image classification tasks. Transformer-based
architectures, such as Vision Transformers (ViTs), replace convolutional layers with self-
attention mechanisms, enabling the model to capture global dependencies in images. You can
investigate the performance and scalability of these architectures compared to traditional CNNs.

Meta-Learning and Few-Shot Learning: Meta-learning approaches aim to enhance the ability
of CNNs to learn from a few labeled examples by leveraging prior knowledge learned from
similar tasks or datasets. Few-shot learning techniques, such as meta-learning, metric learning, or
generative modeling, enable CNNs to generalize to new classes with limited training data. You
can explore the advancements in meta-learning and few-shot learning for image classification
and compare their performance with traditional CNN models.
AutoML and Neural Architecture Search: Automated Machine Learning (AutoML)
techniques, specifically Neural Architecture Search (NAS), have gained attention for
automatically discovering optimal CNN architectures for image classification. NAS algorithms
leverage reinforcement learning, evolutionary algorithms, or gradient-based optimization to
search for architectures with improved performance. You can discuss the progress in AutoML
and NAS and evaluate their effectiveness in discovering superior CNN architectures.

Explain ability and Interpretability: As CNNs become more complex, understanding the
decision-making process of these models becomes crucial. Future advancements in CNNs for
image classification should focus on improving interpretability and explainability. You can
explore methods like attention visualization, saliency maps, or class activation maps that provide
insights into which regions of an image contribute most to the classification decision.

Robustness and Adversarial Defense: CNNs are susceptible to adversarial attacks, where
subtle perturbations to input images can lead to misclassification. Future advancements in CNN
architectures should address the robustness and security concerns by incorporating defenses
against adversarial attacks. You can discuss different defense mechanisms and compare their
effectiveness in improving the robustness of CNN models.

Further Discussions
We will go into more detail about the proposed EvoCNN method's fitness evaluation, weights-
related parameters, and architectures' encoding strategies in this paragraph. The experimental
findings are also reviewed, which may offer helpful information about the potential uses of the
suggested EvoCNN approach. Mutation operators serve as the exploration search, or the global
search, whereas crossover operators serve as the exploitation search, or the local search. Since
local and global searches should compliment one another, only properly developing both of them
might significantly boost performance. The commonly employed methods for CNN weight
optimisation are based on the gradient data. The gradient-based optimizers' sensitivity to the
beginning positions of the parameters that need to be optimised is well known. The gradient-
based methods are prone to becoming stuck in local minima without a suitable starting point. It
seems impossible to identify a better starting point for the connection weights using GAs due to
the vast amount of characteristics. As we have seen, a sizable number of factors cannot be
successfully optimised or efficiently stored into the chromosomes. An indirect encoding strategy
is used in the proposed EvoCNN technique, which simply encodes the means and standard
derivations of the weights in each layer. The final classification accuracy is frequently taken into
account by methods now in use to find CNN architectures together with an individual's fitness.
The training method normally involves several additional epochs, which takes a long time to get
a final classification accuracy.

Summary:

This article provides an overview of Convolutional Neural Networks (CNNs) for image
classification. It begins by highlighting the importance of image classification in various
domains and the liitations of traditional feature selection approaches. Deep learning, particularly
CNNs, is introduced as a solution to address these limitations.

The article explains that CNNs excel at learning hierarchical representations of images by
utilizing convolutional layers and pooling layers to extract features at different levels of
abstraction. CNNs offer translation invariance, allowing them to recognize patterns regardless of
their location in an image. They are also data-efficient, requiring fewer training examples due to
their ability to capture relevant features and generalize to unseen data.

Transfer learning is emphasized as a key capability of CNNs, enabling them to leverage pre-
trained models trained on large datasets and fine-tune them for specific image classification
tasks. This reduces the amount of training data required and improves classification performance.

Scalability is another advantage of CNNs, as they can be adjusted by adding or removing layers,
changing the number of filters, and modifying the size of filters used in convolutional layers.
This flexibility makes CNNs suitable for various image classification tasks, from simple
classification to complex tasks like object detection and segmentation.

The article outlines the methodology for evaluating CNN performance, which involves training
networks on static and real-time video streams, performing transfer learning, and testing
accuracy. It mentions different types of CNN layers, including input layers, convolution layers,
pooling layers, rectified linear unit (ReLU) layers, and fully connected layers.

Several models and architectures are discussed, such as AlexNet, GoogLeNet, and ResNet50.
The article compares their performance and introduces advancements in CNNs, including
attention mechanisms, transformer-based architectures, meta-learning, AutoML, and neural
architecture search. It also emphasizes the need for explainability and interpretability in CNNs,
as well as robustness against adversarial attacks.

Conclusion

In order to autonomously evolve the architectures and weights of CNNs for image classification
challenges, a novel evolutionary technique is being developed in this study. By putting forth a
new representation for weight initialization strategy, a new encoding scheme for variable-length
chromosomes, a new genetic operator for chromosomes with different lengths, a slacked binary
tournament selection for choosing promising individuals, and an effective fitness evaluation
method to speed up evolution, this goal has been successfully attained. Understanding deep
learning is important, and it is useful since training time is limited. Future study will improve our
system by incorporating evolutionary algorithms to address the classification feature extraction
challenge and reduce the number of parameters required for this operation.

References

[1]. Redmon J, and Angelova A, “Real-time grasp detection using convolutional neural networks”,
IEEE International Conference on Robotics and Automation, pp. 1316–1322, 2015.
[2]. Jasmin Praful Bharadiya https://journaljerr.com/index.php/JERR/article/view/858
[3]. Hang Chang, Cheng Zhong, Ju Han, Jian-Hua Mao, “Unsupervised Transfer Learning via Multi-
Scale Convolutional Sparse Coding for Biomedical Application.” IEEE Transactions on Pattern
Analysis and Machine Intelligence, 23 janvier 2017.
[4]. Jasmin Praful Bharadiya https://journalajorr.com/index.php/AJORR/article/view/164
[5]. Zhou, X., Yu, K., Zhang, T., & Huang, T. “Image classi¿cation using super-vector coding of local
image descriptors.” In ECCV,2010.
[6]. van de Sande, K. E. A., Gevers, T., and Snoek, C. G. M, “Evaluating color descriptors for object
and scene recognition”, IEEE Transactions on Pattern Analysisand Machine Intelligence.” 1582–
1596. 2010.
[7]. Howard, A. , “Some improvements on deep convolutional neural network based image
classi¿cation.” ICLR, 2014.
[8]. Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L., “ImageNet: A large-scale
hierarchical image database.” In CVPR, 2009.
[9]. Ahonen, T., Hadid, A., and Pietikinen, “M. Face description with local binary patterns:
Application to face recognition.” Pattern Analysis and Machine Intelligence, 2037–2041. 2016.