Graph-based multi-source domain adaptation with contrastive and collaborative learning for image deraining

Image deraining networks training on large-scale synthetic datasets have achieved good performance. However, when using these networks on unknown real-world scenes, a performance degradation will occurs due to the domain gap between synthetic and real rains. To alleviate this issue, this paper proposes a Graph-based Multi- source Domain Adaptive Deraining Network (GMDA-DN) with a joint contrastive and collaborative learning scheme. The proposed pipeline consists of the domain-shared image translation module and domain-specific image deraining modules. Firstly, considering the morphology correlations between synthetic and real rain streaks, we develop a Cross-graph Transformation Block (CTB) in the image translation module to produce realistic rain layers via the feature affine transformation guided by real rain information. Secondly, since similar rain streaks exist in multi-scale representations over both synthetic and real-world data. We propose a plug-and-play Scale-aware Graph Reasoning Block (SGRB), which achieves effective cross-scale information complementation and enhancement. Furthermore, to bridge multi-source domains and the target domain, a contrastive learning scheme is designed to push the predicted domain-specific rain layers far away as well as pull corresponding backgrounds closer at the pixel level, and a collaborative learning scheme is presented to align derained results across domains at the feature level. Extensive experiments demonstrate the proposed pipeline's superiority on various synthetic and real-world datasets. For example, on Rain100L and Rain100H, our pipeline outperforms state-of-the-art image deraining methods by 1.5 and 0.5 in PSNR. Codes will be available at https://github.com/wangpengyu0829/.