Multi-Scale Edge-Guided Context Aggregation Network for Single Image Deraining
Jun Wang, Huiyuan Zuo, Zaiyu Pan, Shuyu Han · 2023
DCNNs have shown remarkable results in the task of image deraining. However, many existing methods for single image deraining do not take into account the restoration of edge textures in the image. Some approaches use a backbone network to handle both image deraining and detail restoration. To overcome these limitations, We created a dual-branch structure named Multi-Scale Edge-Guided Context Aggregation Network (MSEGCA-Net) that utilizes a coarse-to-fine approach. This allows us to attain excellent deraining performance while maintaining image resolution. Firstly, we present a main image deraining branch to obtain coarse deraining information from the rainy image, as well as an auxiliary edge texture detection branch to obtain edge texture information. Then, different from previous methods relying on direct guidance from the edge textures, We introduce a new connection block, Edge-Guided Context Aggregation Block (EGCAB), that combines coarse deraining information with edge texture details. This aggregated information guides the primary image deraining branch to produce the ultimate derained image. Our method was evaluated on three datasets: Rain200L, Rain1200, and Rain1400. Both quantitative and qualitative comparisons indicate that our approach excels over current leading deraining methods, showcasing superior robustness against rain and accuracy in preserving image details.