Dual-Branch Attention Dense Residual Network Based on Negative Image for Image Denoising

Currently, most deep denoising networks introduce more parameters, leading to performance degradation, and no research has focused on negative images. Thus, we propose a Dual-Branch Attention Dense Residual Network Based on Negative Image (NDARNet) for image denoising tasks in this paper. NDARNet chooses to widen the network instead of deepening it. It has two parallel branches: one processes the original noisy image, and the other processes its negative image. The negative image inverts pixel values, highlight features that are not very visible in the original noisy image. We also designed an Attention Residual Multi-Scale Block (ARMSB) and an Attention Residual Dilated Block (ARDB) for efficient feature extraction in both branches. We use many long skip connections in the network. Furthermore, our method supports blind denoising. Extensive experiments show that NDARNet achieves favorable PSNR and SSIM results at different noise levels compared to other common and advanced methods.