Research On Image Classification Method Based On GAN

With the rapid development of multimedia technology and the arrival of the era of big data,image classification has become one of the research hotspots in the field of computer vision and pattern recognition.As a combination of generator and discriminator,GAN(Generative Adversarial Networks)is one of the few advanced technologies in the field of deep learning in recent years.In image classification,GAN is gradually applied to supervised and semi supervised image classification tasks with the advantages of extracting real images and generating rich features of images in the training stage.At present,GAN based image classification mainly improves the discriminator to achieve the feature extraction of image classification.The existing image classification accuracy based on GAN still needs to be improved,because: on the one hand,the function of the discriminator in GAN is single,and the ability of feature extraction is weak.On the other hand,it is difficult for GAN to converge,and the nonlinear fitting ability of discriminator is insufficient.To solve the problem of insufficient feature extraction ability of discriminant in GAN,this paper improves the structure of GAN and proposes an auxiliary encoder GAN AEGAN(Auxiliary Encoder GAN)model.By integrating the encoder into the discriminator framework,the encoder and the discriminator share most of the weights,so that the discriminator can extract the main features of the real images and the generated image,so as to improve the single function of the discriminator in the conventional GAN,and make the features extracted by the discriminator more abundant.In addition,an image classification method based on AE-CGAN(Auxiliary Encoder Conditional GAN)is proposed to solve the problem of low quality of generated samples caused by the process of unlabeled generation.To solve the problems of generator's hard to converge and discriminator's nonlinear fitting ability in GAN,a semi supervised image classification method based on DMGAN(Domain Matching GAN)is proposed.Firstly,based on the existing semi supervised GAN,the maximum mean difference loss of true and false sample distribution is introduced into the generator.At this time,the maximum mean difference loss is equivalent to the loss of image content to guide the direction of generator optimization and solve the problem that GAN is difficult to converge;Secondly,in order to increase the nonlinear fitting ability of the discriminator network,the CBAM(Convolutional Block Attention Module)is added to the traditional discriminator structure to learn the important features of the input feature map of the module and improve the classification ability of the discriminator network.The experimental results on MNIST,CIFAR-10,Fashion-MNIST and SVHN datasets show that compared with the traditional image classification method based on GAN,the image classification method based on AEGAN and the semi supervised image classification method based on DMGAN improve the accuracy of image classification effectively.