International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 06 | Jun 2025
p-ISSN: 2395-0072
www.irjet.net
Efficiency of Convolutional Neural Networks in Object Recognition Piyush N. Chavhan Third Year, Department of Artificial Intelligence and Machine Learning, Progressive Education Society's Modern College of Engineering, Pune, (An Autonomous Institution Affiliated to Savitribai Phule Pune University) Shivaji Nagar, Pune-411005, Maharashtra, INDIA ---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Object recognition, a fundamental challenge in
However, CNN architecture makes all the things, it gets features and minimizes overfitting. CNNs use convolutional layers as their fundamental building blocks, where patterns in images (such as edges, textures, and corners) are extracted by applying filters[10]. These properties are used to recognize more advanced patterns in the deeper features of the architecture. The CNN gets something called transformation invariance to translate centric features, meaning the convolution operation needs no idea where what feature is in the picture. This invariance is important for the task of object recognition, where the same object can appear in different images at different scales, positions and angles. By stacking layers of convolution over each other the network can progressively learn more strict and complex features[12].
computer vision, demands accurate feature extraction and pattern recognition to classify and analyze visual data. Traditional machine learning methods often rely on manual feature engineering, which is time-consuming and prone to inaccuracies. This study addresses these limitations by exploring the implementation of Convolutional Neural Networks (CNNs), which automate feature extraction and have significantly advanced the field. Focusing on practical applications, the research highlights techniques like data augmentation, dropout layers, and regularization to mitigate overfitting, while ReLU activation and batch normalization tackle issues like vanishing gradients. By analyzing real-world implementations in tasks such as classification, detection, and segmentation, the research underscores the transformative impact of CNNs and identifies avenues for their broader application in fields like robotics, healthcare, and autonomous systems.
Despite these impressive capabilities of deep neural networks, including CNNs, issues regarding their computation power become the potential drawback of their performance. With such a large number of trainable parameters and operations, it is computationally expensive to train. This issue is generally resolved using ReLU (Rectified Linear Unit) activation functions, which are used more often in CNNs[7]. One likely reason is that the ReLU activation function adds non-linear characteristics, allowing the network to learn more complicated shapes, plus it is computationally efficient since it cleans negative values by zeroing them. Moreover, this activation function overcomes issues like the vanishing gradient problem, which were encountered in older functions such as sigmoid and tanh[5][7][23]. Combined with pooling layers that reduce feature maps through downsampling (such as max pooling), the network reduces dimensionality so that the model has to use fewer parameters to operate on large images[1][2][17].
Key Words: Object recognition, computer vision, feature extraction, pattern recognition, machine learning, Convolutional Neural Networks , data augmentation, regularization.
1.INTRODUCTION Without proper training, AI will struggle to identify common objects. The traditional object recognition algorithms were mostly based on hand crafted features and shallow machine learning techniques. Nevertheless, the development of deep learning approaches, and especially convolutional neural networks (CNNs), has transformed the domain[4]. CNNs do the automatic hierarchical feature learning of the image pixel and thus no manual feature engineering is required. Designed to handle data, for instance, the availability of large-scale labeled datasets like ImageNet[2] and Fashion MNIST[1016] have fueled a revolution in CNNs that enable CNNs to learn rich hierarchical representations from complex objects. These deep learning models have proven so successful that the state-of-the-art in object recognition has improved dramatically leading to systems that are not only more accurate but also become useful to a larger number of applications.
Β© 2025, IRJET
|
Impact Factor value: 8.315
Another important technique that helps to enhance CNN performance, especially when you have less data in datasets is Data augmentation[19][12]. Data augmentation is a technique that improves the generalization of models by applying random transformations such as rotations, scaling, and translations to the input images, creating a more diverse set of input images. This approach minimizes overfitting, due to its exposure to a variety of image instances in every epoch. To apply classification via
|
ISO 9001:2008 Certified Journal
|
Page 696