Abstract:
Addressing the common issue of detail information loss during image processing, which impacts the overall information modeling of images, this paper proposes an image deraining network based on a hybrid attention mechanism and a gating mechanism. Initially, the hybrid attention module employs axial attention to aggregate information at the pixel level along the horizontal and vertical axes of the image, while enhancing global modeling capabilities through cross-channel attention. Subsequently, the fused attention results in feature information. Furthermore, this paper designs a feedforward network incorporating inverted residual blocks based on a gating mechanism. This feedforward network extracts and filters local information features through the combination of inverted residual blocks and the gating mechanism. Experimental results on multiple public datasets demonstrate that the proposed network can effectively remove rain streaks from images, enhancing image clarity and visual effects. Compared to the comparative networks mentioned in the paper, this network achieves a 0.81 dB improvement in PSNR, exhibiting a significant performance advantage.