Enhancing the noise robustness of sparse-form patches for image denoising

As an important prior for distinguishing textures and noise, the sparsity of patches from clear images has attracted much attention in denoising tasks. However, few studies have discussed whether sparse-form patches are robust to noise, which is a key factor for denoising methods to effectively leverage sparse priors. In this paper, we revisit the importance of noise robustness in sparse-form patches used in denoising methods and propose a novel denoising framework, named FeaPD, which can effectively reduce the adverse interference of noise on sparse-form patches. Specifically, the expected patch log likelihood (EPLL) and UNet are taken as our baseline to verify the effectiveness of this framework, resulting in two new denoising methods: FeaEPLL and FeaUNet. First, sparse-form patches are constructed by applying PN on a mean domain with reduced noise intensity, which offers greater robustness to image noise than applying PN directly on the noisy image. Then, matrix-based complete convolution is incorporated for lossless decomposition, and a cross-domain joint probability model is used to enhance texture representation precision by sparse-form patches in the feature domain. Finally, this denoising framework was explicitly validated on the FeaEPLL and further applied to design the advanced FeaUNet. Experiments on synthetic noise using the public datasets Set12, BSD68, Kodak24, McMaster, and Urban100 demonstrate that our methods achieve superior performance with low computational overhead, while also showing competitive results on real-world noise in the SIDD dataset. The code and all pretrained models are available at https://github.com/Xin-Ge/Feature-Patch-Denoiser.