A Semi Supervised Image Rain Removal Network Based on Gaussian Mixture Model

At present, the performance of the deep learningbased image rain removal algorithm for the real rain map in the real world is not good, mostly due to the certain data differences between the synthetic rain image and the real rain image, which leads to the fact that the supervised image rain pattern removal algorithm trained by the deep learning algorithm on the synthetic dataset cannot give full play to its advantages. In addition, there is a very tricky and time-consuming problem in the collection of datasets in the field of image rain removal, and it is difficult to collect real rain maps and their corresponding clean background images to form data pairs at the same time in real life. In order to solve the over-dependence of a single image rain removal algorithm on the synthetic dataset, a single real rain map and paired synthetic rain maps are used as the training atlas, and semisupervised learning is used to train the rain removal algorithm model, which has become one of the hot spots in the research of single image or video rain removal. Based on the previous research, this paper does the research on single-image rain removal based on semi-supervised learning, mainly to solve the over-dependence of the current network model on synthetic rain maps and to provide new ideas for improving the generalization of the image rain stripe model on real rain images. In this paper, we propose a semi-supervised image removal rain streak algorithm based on Gaussian mixture model. We combine the Gaussian mixture model and the multi-stage network to form a deraining model with a semi-supervised learning mechanism. Among them, the Gaussian mixture model is used to model the real rain patterns, the multi-stage network is used to realize the recovery of the background image and the extraction of rain pattern information, and the KL divergence is used to realize the transfer learning of the synthetic rain patterns and the real rain patterns. Based on the Gaussian mixture model, the rainremoving network trained by this semi-supervised mode effectively uses and integrates the advantages of supervised learning and unsupervised learning, and trains the two parts at the same time, so as to reduce the network's dependence on synthetic network images and improve the generalization ability of the network on real-world rain-containing images.