Research on Image Classification of Lightweight Convolutional Neural Network

Abstract

Convolutional neural networks are widely used in related tasks in the image field. However, in order to achieve high accuracy, modern convolutional neural networks often contain a large number of hidden layers and training parameters, resulting in a substantial increase in computing costs. Training and deployment of convolutional neural network models require a large amount of computing and storage resources, which is difficult to apply to smart mobile devices, embedded devices and IoT devices. In order to make the network more efficient, the existing lightweight convolutional neural networks often have shallow depth and limited feature extraction capabilities, and there is still a large room for improvement in accuracy. In response to these problems, a lightweight convolution unit is designed to replace the traditional convolution operation in this paper, and a lightweight convolutional neural network LCUNet based on the VGG-16, VGG-19 and ResNet network structure are designed. These networks are used to do image classification experiments on CIFAR-10 dataset. The experimental results show that the lightweight convolutional neural networks designed in this paper are better than the current efficient lightweight convolutional neural network MobileNet series and ShuffleNet series, and the amount of parameters and calculations are more less.