Research On The Architecture Of Lightweight Convolutional Neural Networks

With the recovery of artificial intelligence,computer vision has been developing at high speed in recent years.As one of the fundamental models in computer vision,convolutional neural networks are the cornerstone of numerous researches and applications in this field.The performance of convolutional neural networks has a direct impact on the upper bound of many other computer vision tasks such as object detection,semantic segmentation,face recognition,and visual question answer.The improvement of convolutional neural networks could lead to the progress of all related computer vision systems.Convolutional neural networks are one of the research highlights in recent years.After years of development,the research of convolutional neural networks has converted from large models with high accuracy to lightweight models which are more suitable for practical applications.The aim of lightweight convolutional neural networks is to keep comparable accuracy with large models,and at the same time reduce the size of models,decrease training and inference time,and even make the models run with embedded devices.With lightweight convolutional neural networks,more researches in computer vision could apply to industrial products and services.This paper focuses on the construction of lightweight convolutional neural networks from the perspective of decomposing convolutions.We argue that decomposing convolutions is the main idea in the progress of lightweight convolutional neural networks.We make a new interpretation of some important lightweight convolutional neural networks from the perspective of decoupling convolutions.Depthwise convolution is one of the primary modules of lightweight convolutional neural networks.Depth-wise convolution decouples the study of spatial correlations from cross-channel correlations.But depth-wise convolution is now the bottleneck of lightweight convolutional neural networks.Shift module and active shift module are efficient alternatives to depth-wise convolution.To improve the limited expressive abilities of shift module and active shift module,we propose a new component of lightweight convolutional neural networks called multi-active-shift module.The main work of this paper could be summarized in the following three aspects:First,we make a new interpretation of the progress of lightweight convolutional neural networks from the perspective of decoupling spatial correlations and crosschannel correlations.Second,we introduce two alternatives to depth-wise convolution,shift module,and active shift module.We prove that shift module is equivalent with standard convolution with sparse kernels and interpret shift module from the perspective of decomposing and re-integrating convolution.Third,we propose a new component of lightweight convolutional neural networks multi-active-shift module to improve the expressive abilities of shift module and active shift module.We use multi-active-shift module to construct a new light-weight convolutional neural network called MASNet,and valid its superiority in speed and accuracy on CIFAR10/100 and ImageNet 2012 datasets.