Center pixel weight estimation for non-local means filtering using local James-Stein estimator with bounded self-weights

Assigning appropriate center pixel weights (CPW) in nonlocal means (NLM) filter is an important issue to affect the quality of filtered images. Using local James-Stein (LJS) type CPW yielded superior peak signal-to-noise ratio (PSNR) over using other existing methods of determining the contribution of center pixels in NLM. However, the original LJS CPW method assumed no upper bound for self-weights implicitly, it may yield excessively large self-weights and visual artifacts. We propose a method, called bounded self-weights LJS (BLJS), to incorporate bounded self-weights into LJS such that this new estimator is dominating partially. Our proposed method was evaluated using a patient MR image with 3 levels of additive Gaussian noise. For fixed local area, the proposed BLJS yielded lower variances than the original LJS for almost all bias levels. BLJS also achieved PSNR comparable to or better than LJS. Visual image quality assessment showed that BLJS produced less localized visual artifacts than LJS.