The Optimization Of Self-supervised Generative Adversarial Nets

Since being proposed in 2014,GAN has achieved good results in computer vision-related fields such as image generation,image style conversion,and image missing completion,as well as speech synthesis and natural language processing.However,the training of GAN has many difficulties: the disappearance of gradients,the collapse of patterns,and the discriminator forgetting the classification boundary.At present,natural images are mainly generated by CGAN,and the generation of GAN is controlled by adding constraints.The main disadvantage of this model is the need to label data.However,labeling the data requires a lot of manpower and material resources,which is very difficult,and even some is not feasible.To address this problem,this paper introduces the idea of self-supervised GAN,which combines two popular unsupervised learning techniques,namely adversarial training and self-supervised learning,which narrows the gap between unsupervised learning and supervised learning.However,in the self-supervised GAN,for the rotation of the generated image,the confrontation loss and the rotation loss are antagonistic,which will cause the quality of the generator to slightly affect the discriminator.Therefore,this paper optimizes the generator loss function of the self-supervised GAN,Further enhancing the stability of GAN training.Since the quality of the generated image in the early stage of the self-supervised GAN training is poor,the image is rotated and the features are extracted for detection,and the results obtained are unsatisfactory.Therefore,this paper proposes to pre-process the self-supervised GAN,and experiments prove that the images generated by the model proposed in this paper are more realistic.By optimizing the self-supervised GAN,the problem that the discriminator forgets the classification boundary is relieved to a certain extent,and the dependence of the discriminator's representation on the generator output quality is reduced,thereby ensuring that the model can be trained more stably.However,the problems of the disappearance of gradients and the collapse of modes of the original GAN have not been solved.It can also be seen in the experiment that the diversity of the image is not well guaranteed.The main reason for these two problems in the original GAN is to use JS divergence to measure the distance between the real distribution and the generated distribution.The JS divergence is a constant when there is no intersection of the two distributions,which causes the gradient to disappear.Therefore,in this paper,the Wasserstein distance is used to replace the JS divergence.Even if the two distributions do not have an intersection,it can still measure the distance between them well,and the Wasserstein distance is smooth,which can provide a meaningful gradient in training..In this paper,Wasserstein distance is introduced on the basis of self-supervised GAN,so that the diversity of generated images is guaranteed.