Bijective Multi-mode Deraining on Single Image

Since rain streaks distribute non-uniformly in most real-world rain images, to avoid over-deraining or under-deraining in local rain image, we design a rain-density distribution dependent rain model. Base on the model, we propose a two-stage deraining network to remove rain streaks appropriately. In stage-I, we propose a Bijective Attention Multi-mode Deraining Network(BAMDN), including a Residual Streak-detail Attention Module(RSAM) to get the attention mask of the rain streak, removes rain streaks in patches with the determined rain-density label. In stage-II, we recombine the patches and propose an Attention Global Restoring Network(AGRN) including Attention Network-II(AN-II), which provides the details on the junctions of adjacent deraining patches to further derain and recover the lost details in stage-I. Comprehensive experiments on synthetic datasets and real-world datasets demonstrate the effectiveness of our method.