International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 08 | Aug 2025
p-ISSN: 2395-0072
www.irjet.net
Image Classification Using Modified Convolutional Neural Network Architecture Dr. Zubin Bhaidasna1, Dr. Hetal Bhaidasna2, Kinnari Mishra3 , Sarthvi Parmar4 1 Department of Computer Engineering, GCET, CVM University, V. V. Nagar, Gujarat, India
2 Department of Computer Engineering, PIET-DS, Parul University, Vadodara, Gujarat, India 3 Department of Computer Engineering, PIET-DS, Parul University, Vadodara, Gujarat, India
4 Department of Computer Engineering, PIET-DS, Parul University, Vadodara, Gujarat, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Image classification is an important task in
deep learning models, Convolutional Neural Networks (CNNs) have become the most dominant approach due to their ability to capture spatial hierarchies through convolution, pooling, and non-linear activation functions [2][5]. CNNs eliminate the dependency on handcrafted features and provide an end-to-end learning framework, significantly improving accuracy in a wide range of tasks including object detection, medical imaging, facial recognition, and natural scene understanding [7][11][21].
computer vision, where the goal is to automatically recognize and categorize images into predefined classes. Convolutional Neural Networks (CNNs) are widely used for this purpose because of their ability to learn visual features directly from data. However, conventional CNN models often face challenges such as overfitting, high computational cost, and limited accuracy in complex situations. In this work, we explore modified CNN architectures to enhance performance. The modifications include adjustments in the number of layers, filter sizes, activation functions, and regularization techniques. These improvements allow the model to capture more meaningful patterns in images and reduce classification errors. Experiments on standard benchmark datasets demonstrate that the modified CNN models achieve higher accuracy and efficiency compared to basic CNN architectures. This study shows that refining CNN design can significantly improve image classification, making it more robust and practical for real-world applications.
In the specific context of binary image classification, distinguishing between cats and dogs has been considered a benchmark problem. This dataset, though simple in appearance, presents challenges such as variations in pose, lighting, orientation, and background noise. Researchers have demonstrated that CNNs can achieve impressive accuracy on this problem, making it a standard testbed for evaluating new architectures [4][9]. Modified CNNs, with additional layers, adjusted filter sizes, or the incorporation of regularization methods such as dropout and batch normalization, have shown potential in improving classification performance even further [6][12][23].
Key Words: Image Classification, Convolutional Neural Network (CNN),Deep Learning; Machine Learning, Feature Extraction, Data Augmentation.
Recent studies suggest that enhancing CNN architectures by fine-tuning hyper parameters, experimenting with different activation functions, and incorporating data augmentation techniques leads to superior generalization [8][13]. Data augmentation strategies like flipping, scaling, and rotation expand the training dataset virtually, allowing the network to learn invariant features that improve robustness. Furthermore, optimization algorithms such as Adam, RMSprop, and SGD with momentum have played a crucial role in accelerating convergence and stabilizing training processes [10][15].
1. INTRODUCTION Image classification has emerged as one of the most significant and widely studied tasks in the field of computer vision. It refers to the process of automatically assigning a label to an image based on its visual content. With the rapid growth of digital data, especially images and videos, the demand for accurate and efficient classification techniques has increased dramatically. Traditionally, image classification relied on machine learning techniques such as Support Vector Machines (SVM), k-Nearest Neighbours (kNN), and Random Forests, where handcrafted features like edges, textures, and colour histograms were extracted manually. However, such methods often failed to generalize well across complex datasets due to their limited ability to capture hierarchical features [1][3][23].
While machine learning-based approaches laid the foundation for automated classification, the shift toward deep CNNs has provided a transformative leap in performance. Nevertheless, standard CNN architectures may still suffer from issues such as overfitting, vanishing gradients, and high computational cost when applied to large-scale image datasets [14][18][25]. Therefore, research has focused on modifying CNN models to achieve a balance between accuracy and efficiency. These modifications
The advent of deep learning has revolutionized image classification by enabling models to automatically learn discriminative features directly from raw images. Among
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 411