Non local means image denoising using noise-adaptive SSIM

This paper embeds SSIM in place of the L₂ norm in a one step Non Local Means (NLM) scheme. This is possible thanks to a new form of SSIM that can be formally derived from the classical SSIM using the spreading error analysis. This approach has several advantages over L₂ norm based NLM such as greater robustness to parameters setting, higher performance in terms of PSNR and SSIM, optimal subjective visual quality. In addition, it is possible to show that the cascade of the proposed pure visual approach and a second step based on L2 norm allows us to reach results close (slightly less) to the state of the art (BM3D) in terms of PSNR and SSIM.