Research On Image Deraining Method Based On Generative Adversarial Network

The research purpose of image deraining problem is to remove the rain streaks from the rainy image and restore a clean background image.Rainy images taken outdoors are susceptible to distortion and blurring due to rainfall,which affects the performance of subsequent highlevel computer vision tasks such as object recognition.Therefore,image deraining has important research significance and application value.In thesis,the image deraining methods based on generative adversarial networks are proposed to address the above problems.The main research work and related innovations of the thesis are as follows:In order to solve the problems of weak background feature extraction ability of complex scene rainy images and blurred details in the recovered background image,the image deraining method based on pyramid pooling and conditional generative adversarial network is proposed in thesis.Conditional generative adversarial networks are used as the infrastructure of the overall network.To improve the ability of the model to extract background feature information and reconstruct background images,the Sparse Feature Reactivation Dense Nets and the Efficient Channel Attention are added to the generative network structure.The Pyramid pooling module is designed to make full use of the global information of the features in the generative network and improve the detail reconstruction accuracy.The discriminative network structure is designed with global-local dual discriminative approach to keep the overall style and local details of the generated images consistent.Through relevant comparison experiments,it is verified that the method not only has good visual effect of rain removal,but also the PSNR and SSIM indexes are improved,which can effectively solve the image deraining problem in complex scenes.Most current image deraining methods overly rely on supervised training using synthesized rainy images,and the data distribution of synthetic rainy images differs significantly from that of the real rainy images,leading to poor generalization of the model in the real rainy images.To solve the problem,the image deraining based method on semisupervised learning and cycle generative adversarial networks is proposed in the thesis.The method uses both real and synthetic rainy images for semi-supervised training.Cycle generative adversarial networks are used as the infrastructure of the overall image deraining network model.The unsupervised training of the model on the real rainy images is constrained by introducing a Gaussian fuzzy prior.To enable stable training of the network model,the discriminative network is designed with a multi-scale structure and incorporates a spectral normalization operation.Compared with the existing semi-supervised image deraining methods,the method has better deraining visualization and generalization on the real rainy images.