Research on lightweight convolutional neural network for mobile devices

• 1、Intelligence Science and Technology Center of BUPT，Beijing 100876

Abstract：In the mobile environment, the deep neural network based on traditional convolution has the problem of large amounts of parameters and calculation. In order to solve this problem, a new lightweight module is proposed. Based on the idea that the feature maps of complex network are redundant, the input feature maps are compressed by group convolution to obtain meta-features, and then the information between groups is fused on the basis of the meta-feature maps, which simplifies the network while ensuring the effective communication between groups. Experiments were conducted on the resnet-18 network based on the CIFAR-10 dataset. The improved network\'s accuracy only lost 0.9%, but the amount of parameters decreased by 14 times, and the amount of calculation decreased by 13 times. Experiments show that this module supports to replace traditional convolution in the original deep neural network as a plug-and-play way, which greatly reduces the amount of network\'s parameter and calculation while maintaining the network accuracy.

Keywords： deep learning convolution neural network lightweight network group convolution