DG-RainDiff: Depth-Guided Dynamic Message Passing Diffusion Model for Mixture of Rain Removal

Abstract

Real-world rain is a mixture of rain streaks and rainy haze. It poses a great challenge for current deraining techniques due to two reasons. First, existing methods consider rain streaks removal and rainy haze removal as separate processes. Second, they are limited in insufficient modeling ability to learn the mapping from the mixture of rain to clean images. To address these issues, we propose a novel Depth-guided Dynamic Message Passing Diffusion Model for the mixture of rain removal, called DG-RainDiff. DG-RainDiff is a joint learning paradigm that integrates depth estimation and image deraining in a diffusion framework. It takes full advantages of additional guidance from depth information and powerful generation ability of diffusion models to significantly improve the capacity in the mixture of rain removal. Furthermore, we also explore the importance of contextual information in image rain removal tasks and introduce a novel dynamic message passing module (DGMP) that contains convolution with offsets to enhance the ability of DG-RainDiff to obtain rich contextual information, thereby achieving a more vivid restoration of occluded pixels. Extensive experiments on both synthetic and real-world data show that DG-RainDiff quantitatively and qualitatively outperforms fifteen state-of-the-art methods.