Fuzzy-Clustering Attentive Network With Context-Aware Enhancement for Image Denoising

Image denoising is a long-standing inverse problem in computer vision. Despite recent progress, existing methods still struggle to balance precise feature focusing with structural fidelity under complex noise, largely due to (i) the winner-takes-all bias of probability-normalized soft attention and (ii) isotropic context fusion that blurs oriented structures. We propose the Context-aware Enhanced Fuzzy Clustering Attention Network (CEFCA-Net) that couples a Fuzzy-clustering Attention (FcA) mechanism with a Context-aware Enhancement (CaE) mechanism. FcA replaces rigid probability normalization with learned fuzzy memberships to multiple trainable cluster centers, enabling parallel, non-competitive multi-feature focusing and better detail retention. CaE adopts an anisotropic design with parallel square, horizontal, and vertical convolutional branches and adaptive fusion, explicitly capturing localized textures and long-range directional dependencies to preserve edges and line structures. The resulting select-and-enhance synergy between FcA and CaE achieves effective noise suppression while maintaining structural fidelity and favorable efficiency. Extensive experiments on synthetic and real-world noisy image benchmarks show that CEFCA-Net outperforms most state-of-the-art methods while operating at markedly lower computational cost, and further analyses indicate that it degrades gracefully across a wide range of noise levels and remains effective under non-Gaussian noise, demonstrating favorable noise adaptability, detail preservation, and computational efficiency.