Multi-stage image rain removal network based on Swin Transformer

• 1、School of Computer Science, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Various vision tasks such as object detection are generally processed based on clear images when applied, and degraded images will reduce the accuracy of these vision algorithms. Rainy days will block and blur the captured images, which will degrade the quality of the captured images. Therefore, image rain removal is currently a research hotspot. However, the existing rain removal algorithms still have the problems of rain streak residue and blurred background texture. Therefore, a multi-stage image raining algorithm based on Swin Transformer is proposed in this paper. The algorithm can effectively pay attention to the global information of the image, make full use of positive and negative samples, and gradually obtain better rain removal effect. Specifically, the network is based on the Encoder-Decoder framework, where the Encoder is partly based on the Swin Transformer. In addition, the network introduces contrastive loss, and treats the rain image and the clean image as negative samples and positive samples respectively, to ensure that the restored image is close to the clear picture and away from the rain image in the feature space. The network has two stages, using a recursive layer to connect, gradually removing rain streaks and restoring a clean background. Through extensive experiments on several challenging datasets, it is proved that the proposed method has better performance compared to existing algorithms.????

Keywords： rain removal Swin Transformer multi-stage contrastive loss