| In recent years,in the field of computer vision,deep learning has promoted the rapid development of many technologies,such as face recognition and autonomous driving.However,in practical applications,deep learning models require many image samples for training,and it is often very difficult to obtain image training samples due to privacy,cost and other issues.Generative Adversarial Nets(GAN)can generate many realistic images to supplement training samples,which makes a contribution to solving the problem of difficult to obtain image training samples.As a variant of GAN,Wasserstein GAN with Consistency Term(CT-GAN)improves the stability of the model during training.However,CT-GAN still has some instability problems during training,such as i nitial fluctuation and difficulty in rapid convergence.In order to further improve the quality of images generated by CT-GAN model,this paper analyzes the instability problems of CT-GAN model,and summarizes the causes of its instability into two aspects :(1)The discriminator is improper in dealing with the 1-Lipschitz constraint;(2)Lack of constraints on generator.For the first reason,an Asymmetric two-sided Penalty Term is constructed based on the gradient penalty method to improve the objective function of CT-GAN discriminator.On this basis,WGAN-AP model is established in this paper.The model adopts different penalty methods for different gradient norm in a specific region and pays attention to generated sample edge region,which realizes a more relaxed and broader constraint.Experiments on Toys,MNIST,CIFAR-10 and Image Net datasets show that the WGAN-AP has smaller fluctuation in the early stage,can converge to a stable state faster and improve the clarity and diversity of the generated ima ges.For the second reason,this paper proposes different optimization schemes for the proposed WGAN-AP under supervised learning and unsupervised learning modes based on regularization technology.In the supervised learning mode,the weight matrix of Linear,Residual Block and Conv layers in the generator network is spectral normalized,and the WGAN-AP +SN is established.In the unsupervised learning mode,a constraint term guided by the discriminator gradient is added into the parameter iteration process of the generator,and the WGAN-AP +GC is established.Through experiments on MNIST,CIFAR-10 and Image Net datasets,this paper shows that the optimization scheme can make the WGAN-AP more stable and generate higher quality image samples in both supervised learning and unsupervised learning modes.To sum up,in order to relieve the pressure of insufficient image training samples of deep learning model,this paper studies the image generation model CT-GAN.Finally,this paper improves the training stability o f CT-GAN based on gradient penalty and regularization techniques,and improves the quality of image generated by the model. |
