A two-stage network with wavelet transformation for single-image deraining

Image deraining is still a vital and challenging low-level computer vision task, the purpose of which is to restore rain-free images from images degraded by rain streaks. Recently, many convolutional neural network (CNN)-based methods have made significant progress for the task of image deraining. However, there are still two issues with these methods: The rain streaks in heavy rain images cannot be effectively removed, and the high-quality images with clear details cannot be reconstructed. To solve the first issue, we adopt a two-stage structure to gradually remove rain streaks and use dilated convolution in the first stage network to rapidly enlarge the receptive field to capture the spatial characteristics of heavy rain streaks. For the latter issue, we introduced a structure-preserving network (SPN) without any up- and down-sampling and designed a feature extraction module based on wavelet transform in SPN to help SPN restore clear high-frequency details. In addition, we also designed a feature filter (FF) and multi-level feature fusion module (MLFFM), so that the valuable features in the first stage can be fully utilized in the second stage. Extensive experiments on six synthetic datasets and one real-world dataset indicate that our method can achieve excellent performance, compared with the current state-of-the-art methods. To further demonstrate the practical applicability of the proposed method, we use it as a semantic segmentation preprocessing step and display the semantic segmentation results to verify the effectiveness of our deraining method in downstream vision tasks.

Automated summary

Image deraining is an important and challenging computer vision task. This study proposes a two-stage method to effectively remove rain streaks and reconstruct high-quality images. The method utilizes a structure-preserving network and a feature extraction module to enhance the details. Extensive experiments show excellent performance on synthetic and real-world datasets, and the method proves to be effective in downstream vision tasks.