A COMPARATIVE STUDY OF NEURAL NETWORK ARCHITECTURES FOR IMAGE RECOGNITION TASKS

Abstract

Computer vision, medical image analysis, and autonomous vehicles are just a few of the many areas that have found success using Convolutional Neural Networks (CNNs), a potent tool for picture recognition. Choosing the correct CNN architecture and training method, however, can be difficult when working with huge datasets. The paper compares different convolutional neural network (CNN) designs and training strategies to determine which is the most effective for photo recognition. A variety of designs and training techniques for convolutional neural networks (CNNs) for image recognition are covered in the literature review, which covers earlier research on the subject. This paper outlines the pros and cons of popular convolutional neural network (CNN) designs including LeNet, AlexNet, VGG, and ResNet. Common training methods are also discussed in the literature, such as SGD, Adam, and BN. The research relied on the CIFAR-10 dataset, which included sixty thousand 10-category colour images at a resolution of thirty-two by thirty-two pixels. As a first step in the data preparation procedure, we normalised the pixel values and added some random rotations and flips to the training set. Using three distinct training methods—SGD, Adam, and SGD with BN—the researchers constructed and trained seven distinct CNN architectures, namely LeNet, AlexNet, VGG-16, VGG-19, ResNet-50, ResNet-101, and ResNet-152. With a 94.7% success rate, ResNet-152 proved to be the best architecture for the CIFAR-10 dataset. ResNet-101 and VGG-19 were just behind, both reaching 93.7% accuracy. With 152 layers, ResNet-152 outperformed VGG-19, which has only 19 layers, demonstrating that deeper networks beat shallower ones. Improved performance of the CNN architectures, leading to faster convergence and higher accuracy, was achieved by incorporating BN into the SGD training technique. This study sheds light on the relative merits of various convolutional neural network (CNN) designs and training methods for picture recognition. High accuracy in image recognition tasks is achieved by carefully choosing the CNN architecture and training technique, as shown by the findings. Adding BN to SGD enhanced performance, highlighting the importance of the training strategy in the study. These results have real-world consequences since they may guide future CNN designs by experts in the area of image recognition. Research in the future might look into how well these CNN designs and training methods work with other datasets, as well as investigate other cutting-edge approaches like adversarial training and transfer learning. Computer vision, medical image processing, and autonomous cars are just a few of the many potential applications of convolutional neural networks (CNNs) that could be made possible by this type of research.