Abstract: In recent years,with the development of unmanned driving,intelligent UAV and mobile Internet,the demand for lightweight neural network from low-power,low-cost mobile and embedded platforms is increasingly urgent.Based on the idea of deformable convolution and depthwise separable convolution,this paper presents a low-cost deformable convolution,which has the advantages of high-efficiency feature extraction ability of deformable convolution and low computational complexity of depthwise separable convolution.In addition,on the basis of applying low-cost deformable convolution and combining with the method of model structure compression,4 lightweight methods of MobileNet network are designed.Experiments on Caltech256,CIFAR100 and CIFAR10 datasets demonstrate that low-cost deformable convolution can effectively improve the classification accuracy of lightweight networks without significant increase in computational effort.Besides,the accuracy of the MobileNet network can be improved by 0.4%~1% by combining the 4 MobileNet re-lightening methods in this paper,while the network computing load can be reduced by 5% ~ 15%,which significantly improves the performance of the lightweight network and better meets the practical needs of low power consumption and low computing power.It has very important practical significance for the advancement of intelligence in the field of mobile and embedded platforms.

Key words: Low-cost deformable convolution, Deformable convolution, Depthwise separable convolution, MobileNet, Lightweight, Image classification