ECA-MSNet: A Multi-Scale Residual U-Net with Efficient Channel Attention for Real-World Image Denoising

Real‑world photographs contain complex, sensor‑dependent noise that simultaneously obscures subtle high‑frequency textures and broad contextual cues, making denoising a persistent challenge in low‑level vision. The goal of this study is to devise a single, computationally balanced model that removes such heterogeneous noise while faithfully preserving both fine detail and global structure. We introduce ECA‑MSNet, a dual‑branch convolutional architecture designed around this objective. The Residual Detail Estimation Branch reconstructs delicate textures that are most susceptible to corruption, whereas the Multi‑Scale Feature Restoration Branch—a U‑Net enhanced with Attention‑based Multi‑Scale Residual Blocks and lightweight Efficient Channel Attention (ECA)—captures coarse‑to‑fine contextual information. A Dual Residual Fusion Module adaptively merges the two outputs, and a final Refine Block suppresses residual artifacts, yielding the restored image. Extensive experiments on the SIDD and PolyU real‑noise benchmarks validate the effectiveness of the proposed method. ECA‑MSNet achieves 39.41 dB / 0.9109 SSIM on SIDD and 37.76 dB / 0.9574 SSIM on PolyU, outperforming strong baselines such as DnCNN, FFDNet, CBDNet, and CycleISP. Ablation studies further confirm that each architectural component—dual‑branch design, multi‑scale residual blocks, channel attention, and fusion strategy—contributes measurable gains. These results demonstrate that ECA‑MSNet sets a new state of the art for real‑world image denoising, offering a favorable trade‑off between accuracy and efficiency and providing a versatile foundation for other low‑level vision tasks.