Multi-level wavelet network based on CNN-Transformer hybrid attention for single image deraining

Removing rain streaks from rainy images can improve the accuracy of computer vision applications such as object detection. In order to make full use of the frequency domain analysis characteristics of wavelet and combine the advantages of Convolutional Neural Network (CNN) and Transformer, a Multi-level Wavelet Network Based on CNN-Transformer Hybrid Attention (MWN-CTHA) for single image deraining is proposed. MWN-CTHA obtains multi-scale low-frequency and high-frequency images through multi-level non-separable lifting wavelet transform and uses CNN-Transformer Hybrid Attention Block (CTHAB) to learn global structure and detail information from low-frequency and high-frequency, respectively. CTHAB consists of CA-SA Layer (CSL) and Detail-enhanced Attention Feed-forward Layer (DAFL). CSL uses the non-local modeling ability of self-attention to capture long-range rain streaks and uses convolutional attention to enhance the search ability for local rain streaks, where convolution can assist self-attention to achieve better feature representation. DAFL utilizes Depth-wise Convolutional Layer to supplement detailed features and filters the information of feed-forward layer through Dual-branch Attention. The experimental results on the four synthetic datasets demonstrate that the proposed method achieves higher PSNR and SSIM than the state-of-the-art method DANet, with an improvement of 1.07 dB and 0.0098, respectively. The code is available at https://github.com/fashyon/MWN-CTHA.