A Lightweight Deraining Network Based on Hash Attention

Rain streaks frequently result in severe visual degradation and occlusion of the foreground, significantly hampering the execution of visual tasks in outdoor scenes. Currently, the majority of image deraining methods heavily rely on Convolutional Neural Network (CNN). Nevertheless, the crucial issue of balancing computational cost and deraining performance remains unresolved. Therefore, we propose a deraining network based on hash attention to achieve a balance between rain removal performance and computational complexity. To reduce computational complexity, we design the Hash Transformation Attention Block (HTAB), which enables fast filtering and extraction of input feature information. Additionally, to enhance the quality of feature learning, we propose the Controlled Feature Transformation Block (CFTB), which introduces a feature control mechanism to adjust the detailed complementarity between information flows at multiple levels. The experimental results demonstrate that the proposed network HAD-Net achieved outstanding image deraining performance with relatively low computational cost.