Single image deraining using contrastive perceptual regularization

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

Abstract licence: CC BY-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.