A Dual Branch Attention Network for Multiple Degraded Image Restoration

Abstract

In this paper, we propose a novel dual branch attention network (DBA-Net), which can efficiently restore multiple degradation images. Most methods of image restoration focus on a single degradation factor, such as blur, noise, raindrop, etc. However, these methods frequently fail to real degraded images, because the real images usually contain different types of degradation. Our network backbone consists of two branches, the residual branch and the information distillation branch, which have different receptive fields. We design a gate module to choose useful feature maps from the two branches. Then, an improved multi-channel attention selection module is proposed to allow the network to learn more features of real results. On DIV2K datasets, which contains different degradation types (noise, blur, and compression loss) and diverse degradation levels (mild, moderate and severe), our results on PSNR and SSIM outperform other advanced algorithms. In visual performance, the texture of objects in DBA-Net images is more similar to the original image.