Real image denoising algorithm based on the feature adaptive sharing network

Aiming at insufficient high-frequency information extraction and poor cross-scale feature dynamic adaptability in real-world image denoising, we propose a denoising algorithm based on the Feature Adaptive Sharing Network (FASNet). Built on a single U-shaped backbone, cascaded Multi-scale Feature Dynamic Fusion Modules (MFDMs) run through the whole pipeline and act as the bottleneck for encoder-decoder information matching. Each MFDM consists of a Nonlinear Gating Module (NGM) and a Multi-scale Feature Sharing Module (MFSM), combining dilated and standard convolutions with layer-wise adaptive weight sharing via the maximum receptive field unit, and adopts a dynamic aggregation strategy to enhance feature discriminability. Experiments on the DND and SIDD datasets show FASNet achieves 39.85/39.77 dB PSNR, 0.955/0.958 SSIM and optimal NIQE, outperforming 11 state-of-the-art algorithms. Subjective MOS tests and generalization experiments on RNI15 and real smartphone images further verify its superior visual quality and strong robustness.