Image Recognition Research Based On Anisotropic Dilated Convolution

Image is an important way for human beings to obtain information.With the increasing degree of social informatization,the need for automatic image recognition has become more and more urgent.Because of the rapid development of Internet and mobile Internet,image data in the network shows explosive growth,while the recognition effect of traditional pattern recognition methods on massive image data sets is unsatisfactory.Researchers propose to use Convolutional Neural Network(CNN)for image recognition,in which convolution is the most critical component.Existing convolutions such as ordinary convolution,dilated convolution and deformable convolution have achieved good results in image recognition,but there are still some shortcomings:the range of the receptive field of ordinary convolution is limited by convolution kernels,which is not robust enough;dilated convolution will form a regular receptive field,not enough to cope with changes in the apparent characteristics of objects;the shape of the receptive field of deformable convolution can not be quickly transformed,and can not locate the object accurately.In order to solve the above problems,this paper makes a thorough study of the above convolutions,and independently completes the following work:(1)On the basis of ordinary convolution,dilated convolution and deformable convolution,an anisotropic dilated convolution(ADC)is proposed.The advantage of anisotropic dilated convolution is that it inherits the advantages of ordinary convolution,dilated convolution and deformable convolution.It is a more generalized convolution form.Other convolutions can be considered as a special case of anisotropic dilated convolution.Anisotropic dilated convolution can form irregular receptive field,which can be flexibly deformed according to the location of pixels and its own characteristics.While expanding the range of receptive field rapidly through the dilation rate,it can also make the range of receptive field fit the identified object better through the learning of offset,so it can locate the object accurately.(2)Based on anisotropic dilated convolution,Anisotropic Dilated Convolutional Network(ADCN)is established.Anisotropic dilated convolutional network is a network obtained by using anisotropic dilated convolution to replace ordinary convolution or dilated convolution in baseline network(such as VGG,ResNet,etc.).Compared with baseline network,anisotropic dilated convolutional network can retain effective information and remove invalid information,so as to reduce the loss of spatial information and extract more robust features.In this paper,experiments are carried out on seven data sets(MNIST,CIFAR10,CIFAR100,SVHN,STL-10,LFW,ImageNet)and four benchmark networks(VGG,ResNet,WRN,DRN).Firstly,the feasibility and effectiveness of anisotropic dilated convolution are verified;secondly,the best network layer using anisotropic dilated convolution is determined by experiments;finally,the anisotropic dilated convolutional network is applied to image classification and target location.The experimental results show that the anisotropic dilated convolutional network outperforms the benchmark network and achieves the best results in the image classification tasks of MNIST and STL-10.