C2-Matching: Transformation-Robust Correspondence Matching for Image Restoration

Figure 2. (a) Overview of the proposed C2-Matching. The contrastive correspondence network is designed for transformation-robust correspondence matching. The student contrastive correspondence network takes both the LR input image and HR reference image (the transformed version of the HR input image serves as the HR reference image during training) as input. The descriptors before and after transformations are pushed closer while distances of the irrelevant descriptors are maximized by L' margin. To enable the student LR-HR contrastive correspondence network to perform correspondence matching better on highly textured regions, we embed a teacher-student correlation distillation process to distill the knowledge of the easier HR-HR teacher matching network to the student model by L' kl. (b) Overall pipeline of the restoration network. The correspondences are first computed by the trained student contrastive correspondence network, after which the correspondences are used for subsequent dynamic aggregation module and restoration module.