Comparison Of Conventional and Lightweight Convolutional Neural Networks for Image Classification

Abstract

With the popularity of artificial intelligence, the use of deep learning is more and more extensive, but devices often need good performance to complete this task. However, for mobile devices and some devices with relatively poor performance, deep learning seems to have many obstacles. So, this article explores lightweight neural networks. I used two datasets to explore the problem of using lightweight neural networks to classify images. One is the issue of binary classification, which uses gender data sets. The other is the multi-classification problem, which uses bird data sets. I compared three models, one of which is a relatively complex neural network, and the other two are respectively MobileNet and ShuffleNet.In the first dataset, that is, the binary classification problem, the relatively complex neural network performs well, and the accuracy reaches 0.93. However, the accuracy of MobileNet with pre-training parameters and ShuffleNet without pre-training parameters also reached 0.86 and 0.84 respectively. Although the accuracy is reduced, it also shows that the lightweight neural network can well complete this problem. In the multi-classification problem, the relatively complex network performance has been greatly reduced. The accuracy is only 0.76, which may be due to the limited ability of standard convolution to process information. The accuracy of MobileNet with pre-training parameters is 0.87, and the accuracy of ShuffleNet without pre-training parameters is 0.62. So, the conclusion is that using lightweight neural networks should pay attention to the complexity of the problem, whether to use pre-training parameters and the selection of parameters.