Research On Generative Adversarial Network GP Algorithm Based On Difference And Its Application

Generative Adversarial Networks(GAN)is is a hot and frontier research direction in Deep Learning.The cross-application research of GAN and other fields is also more and more extensive.However,the quality of image generated by the original GAN is low and there are problems of gradient disappearance and mode collapse.In this thesis,we mainly study Wasserstein GAN(WGAN)and the application of CycleGAN in Image-to-image translation.The specific innovations are as follows:(1)A WGAN gradient penalty algorithm based on difference(WGAN-dgp)is proposed.WGANdgp algorithm is based on WGAN.It connects two points in the real data distribution and the generated data distribution.It calculates the gradient penalty term of any two points on this line segment based on difference and adds the maximum and minimum loss function in WGAN to make the continuity constraint of gradient penalty term stronger.The simulation results show that the algorithm can accelerate the convergence,produce better image quality and have better robustness.(2)A CycleGAN algorithm based on dual adversarial and its application in image local style translation are proposed.Traditional style transfer algorithm is prone to change the whole image in the process of processing image local style translation.Based on CycleGAN network structure,a suppressor network is added to suppress the change of images.Firstly,the generator acts on the image to get the image of the corresponding domain.Then the image is refined through the suppressor to get the final image.Experiments show that the new algorithm can make the background change less and the texture clearer in the local style translation.(3)A gradient penalty algorithm based on difference for dual adversarial CycleGAN is proposed.Combining WGAN-dgp algorithm proposed in Chapter 3 with CycleGAN algorithm based on dual adversarial proposed in Chapter 4.Channel attention module and spatial attention module are added into the network of the generator.The simulation results show that the image texture generated by the algorithm in image local style translation is clearer and the generated image distribution is closer to the real data distribution.