When dual contrastive learning meets disentangled features for unpaired image deraining

As the basis work of image processing, rain removal from a single image has always been an important and challenging problem. Due to the lack of real rain images and corresponding clean images, most rain removal networks are trained by synthetic datasets, which makes the output images unsatisfactory in practical applications. In this work, we propose a new feature decoupling network for unsupervised image rain removal. Its purpose is to decompose the rain image into two distinguishable layers: clean image layer and rain layer. In order to fully decouple the features of different attributes, we use contrastive learning to constrain this process. Specifically, the image patch with similarity is pulled together as a positive sample, while the rain layer patch is pushed away as a negative sample. We not only make use of the inherent self-similarity within the sample, but also make use of the mutual exclusion between the two layers, so as to better distinguish the rain layer from the clean image. We implicitly constrain the embedding of different samples in the depth feature space to better promote rainline removal and image restoration. Our method achieves a PSNR of 25.80 on Test100, surpassing other unsupervised methods.