Patch Similarity Based Anisotropic Diffusion for Image Denoising

A patch-similarity-based anisotropic diffusion is presented for image denoising. The traditionally anisotropic diffusion based on the intensity similarity of each single pixel (or gradient information) cannot effectively preserve weak edges and details, such as texture. The non-local means algorithm based on patch similarity can preserve texture details well, since the non-local means utilizes the intensity similarity of neighbor pixels. In the proposed method, the patch similarity is adopted to construct the diffusion function of the anisotropic diffusion for gray and color image denoising, namely the diffusion coefficient of the anisotropic diffusion depends on the patch similarity, not image gradient. Therefore, the diffusion of this method is more effective and accurate than the traditional method, because image patches can represent structure information, such as edge and texture etc, while single pixel cannot represent structure information. Experimental results of gray and color images demonstrate that the proposed method can preserve details better than the traditional anisotropic diffusion, and has not noticeable staircase effect that appears in the traditional anisotropic diffusion. In addition, the time complexity is lower than that of the non-local means algorithm. Brain and cardiac magnetic resonance (MR) images and brain Chinese visible human image denoising experiments also indicate that the proposed method has a promising application.