Review of research on lightweight convolutional neural networks

Abstract

The application of convolutional neural networks is becoming more and more mature, and the use of mobile terminals is becoming more and more widespread. In order to deploy convolutional neural networks to mobile terminals, it is necessary to carry out lightweight design of the network, mainly from the aspects of parameter quantity and calculation quantity. At the same time, the accuracy of the network must be guaranteed. The main idea of lightweight model design is to design a more efficient “network computing method”, mainly for the convolution method. Firstly, several convolution methods are introduced, then six lightweight convolutional neural networks that have been excellent in recent years are described, and the innovations of the model are discussed. Next, the accuracy and parameters of each model on the ImageNet data set are analyzed, and the lightweight techniques of each model are further compared. The importance of group convolution, 1 × 1 convolution, residual unit, and channel information circulation is found. Also combining lightweight design with reinforcement learning and neural network architecture search methods will have better results. Finally, the lightweight convolutional neural network is summarized and prospected.