Investigations on Convolutional Neural Network Based on Image Recognition

Abstract

Many convolutional neural networks have emerged over the years, varying in accuracy, speed, and architecture. From LeNet-5 to GoogleNet, the CNN model developed rapidly, and the architecture of the model became more complex. These models are known for their accuracy on ImageNet. Hence, the theme of this study is to explore and study classical CNN models, because these models are similar in basic structure, but they modify the model in different ways. In addition, this paper selects four classical CNN models for learning and implementation, uses FashionMNIST datasets on four models, and records the performance differences of image classification of specific models, with the intention of studying the comparison of results among different models. The four models are LeNet-5, AlexNet, VGG-16, and GoogleNet. The training was carried out on the same equipment and under the same conditions. The study finds that GoogleNet achieves the best prediction accuracy. LeNet-5 spends the least amount of time on training and forecasting. The GoogleNet and AlexNet models can be considered for practical applications, while the VGG-16 and LeNet-6 are either inefficient due to the long training time of the models, or the accuracy cannot.