Single image deraining using contrastive perceptual regularization

Rain streaks pollute the image captured from outdoor vision system, and single deraining approach based on data-driven has witnessed continuously growing achieved great success. Here, an end-to-end network for is proposed. Firstly, to address limit of convolution neural (CNN) which can only extract local feature, a graph basic block proposed global feature. The consists convolutional (GCN) CNN. GCN module combines spatial coherence computing channel correlation introduced non-local information. While CNN module, attention pixel attention, used earn more weight important features. Secondly, contrastive perceptual regularization adopted enhance loss function, natural restored by utilizing information both positive negative samples with regularization. pulled closer clear pushed farther away rainy image. experiment results several datasets demonstrate that these methods achieve better than previous state-of-art methods.