A Hybrid CNN and Transformer Network for Single Image Deraining

Image degradation caused by rain streaks can seriously affect the accuracy of computer vision algorithms. The performance of the existing deraining algorithms still has room for improvement. In terms of the training dataset, many deraining methods only consider the rain layer. However, rain images are also affected by other factors, such as fog, light. In this work, we first synthesize a rain dataset based on standard optical model, and then a deraining network based on hybrid CNN and Transformer is proposed. Specially, a hybrid attention module is first introduced to generate atmospheric attention maps and transmission maps. Then, a multiscale integration attention module is used to produce rain streak maps. Finally, a refinement module is used to process preliminary deraining images generated by the three attention maps. Extensive experiments on both generated synthetic and real-world datasets have demonstrated the superiority and effectiveness of the proposed method.