An enhanced nonlocal-means algorithm for image denoising

In this paper, we propose an enhanced nonlocal-means (NL-means) algorithm for reduction of Gaussian (or light-tail distributed) noise of a gray-scale image. In the NL-means approach, for each pixel, the method computes a weighted average of all pixels in the image. More weight is given to a pixel if its neighborhood (surrounding pixels) is more similar to that of the pixel being denoised. The neighborhood similarity measure used in NL-means is a L2 weighted norm. We show that using such similarity measure sometimes gives the same weight to neighborhoods with similar and dissimilar edge pattern to the current neighborhood edge pattern. That is while the neighborhood with a dissimilar edge pattern should receive a lower weight (or removed all together) to preserve the original edge content that is an important indication of perceptual image quality. To this aim, we modified the similarity measure to incorporate the edge pattern similarity of the inputs as well. Superior denoising performance as compared to the traditional NL-means is shown in various simulations both visually and in terms PSNR and SSIM (as a good estimate of subjective image quality).