Single image deraining via nonlinear recursive Conv-Transformer

Image deraining is an important foundation of advanced image processing, aiming to reconstruct high-quality clean images from rainy images. In recent years, image deraining algorithms based on convolutional neural networks have become mainstream, and deep learning models based on Transformer have also made significant progress in the field of image restoration. Compared with convolutional neural networks, the image deraining performance of Transformer is slightly inferior. In addition, although excessive parameterization helps improve the generalization performance of Transformer, it can also lead to the network size to be too large to be trained. To address the aforementioned issues, a lightweight image deraining network called the nonlinear recursive Conv-Transformer network is proposed, which not only outperforms the Transformer model but also the convolutional model in terms of performance. Specifically, a dual-branch based on convolutional and Transformer models is first proposed, which integrates the local features extracted by the convolutional model with the global features extracted by the Transformer model; then, a nonlinear projecting block is proposed to implement the constraint recursion and a channel attention module is utilized to fuse multi-branch residual features, which helps in designing lightweight networks. Experiments on a large number of benchmark datasets have demonstrated that the performance of the proposed method is superior to that of state-of-the-art methods, while the computational complexity and parameter quantity are much lower than those of similar methods based on Transformer.