Image Denoising Methods Based On The Non-Local Means

Image will be afffected by the noise inevitably in the process of formation,transmission and record, which makes people unable understand the informationcorrectly from image. Therefore, it has become one of the most important fields inimage denoising and more and more scholars pay attention to it.Buades proposes the nonlocal means (NLM) algorithm based on the study of theanisotropic diffusion, neighborhood filtering and total variation filter. NLM is used inimage denoising and it is possible to obtain a good denoising effect. The basic idea ofnonlocal means filter is to use a large amount of redundant information which is similarwith the neighborhood pixels in the image gray value according to the weightcoefficient to calculate the weighted average of gray estimate for noise pixels. The coreissue of nonlocal means filter is to determine the weighted kernel function. Buades usegaussian weighted Euclidean distance between neighborhood pixels as the weightcoefficient when nonlocal means filter is put forward. The method has better denoisingeffect in relatively flat area of image, but the denoising ability is weak in the rich edgeand texture regions of image. Therefore, using a gaussian weighted distance todetermine the weight coefficient has some limitations.In this paper, we introduce structure similarity into the small image similarity tomeasure based on nonlocal means filter and define a new distance to determine theweight coefficient. At the same time, we also verify the effectiveness of the newalgorithm on the multi-scale. The experiment results show that the denoising effect isimproved and image structure information is preserved better. In addition, we alsointroduce the structure detection into NLM algorithm. We use the results of the edgedetection to adjust similarity measure in NLM algorithm. In order to preserve the edgeinformation of image, the similar content of edge pixels will obtain greater power, andthe dissimilarity contents of the edge pixels will obtain smaller power (or zero). Theexperiment results show that the new algorithm has greater denoising ability than NLMalgorithm and it can preserve edge structure information better, so the visual effect ofimage is better.