Image Denoising Network Based on Multi-Scale Attention Mechanism

Image denoising requires simultaneous suppression of stochastic corruption and preservation of local texture and large-scale structure. This study presents the Multi-scale Attention Denoising Network (MADNet), an encoder–decoder residual architecture that integrates a four-branch Multi-scale Attention Module (MAM) with a Hierarchical Feature Selection Module (HFSM) on the skip pathways. MAM combines a local 1×1 branch, two 3×3 atrous-convolution branches with dilation rates 2 and 4, and a global-average-pooling branch through image-level softmax weighting. HFSM refines encoder features through sequential channel attention, spatial attention, and gated skip propagation. Separate models are trained for additive white Gaussian noise levels σ∈{15,25,50} and evaluated on Set12 and BSD68 in the luminance domain. Under the unified evaluation protocol, MADNet achieves 30.96/0.8748 PSNR/SSIM on Set12 and 29.56/0.8409 on BSD68 at σ=25; at σ=50, it achieves 27.95/0.8098 and 26.67/0.7418, respectively. Component, branch-fusion, dilation/patch-size sensitivity, scale-weight, upsampling, and real-noise analysis quantifies the contribution of each design choice.