Image denoising is a major issue in image processing, and it can affect the subsequent image processing results, ultimately determining the visual effect of images. Besides, when conducting denoising for images, it is required to keep the detailed information of the image as much as possible. With regard to this respect, Research team of Institute of Optics and Electronics proposed a non-local average filtering improved algorithm that keeps image edges and textures.  

When estimating noise pixel, this algorithm adopts information of edges and textures instead of original information of distance to determine the weight of reference pixel in the calculation, while the weights of edges and textures are determined by directional gradient and coefficient of variance. In the selection of reference points of points (datum point) to be processed in the noisy image, given the possibility that the reference point and the datum point are both in the texture zone, it is recommend ed to reduce the filtering strength to save the texture. Thus, texture measurement factor is introduced in this algorithm: the bigger the factor is, the ampler the texture in the pixel zone becomes; meanwhile, to save the image edge information during denoising to the largest extent, edge detection shall be conducted for the original image. According to the result of edge detection, in determining the reference points for a datum point located on the margin of an edge, if one reference point is also margin point, then it shows the resemblance of the two points is likely to be higher than the case when the reference point is in flat zone. Therefore, the resemblance factor of directional gradient in adjacent zones of the two points is added to the weight. Based on the improvements of the above two points, the calculation equation of weight of reference point to the datum point in the improved algorithm is ultimately obtained.  

It is proven by tests that, for processing effects of noisy images with different noise levels, the SSIM appraisal value of new the algorithm is improved by 7.6% compared with that of the original algorithm; through tests on different images, it is shown that the SSIM appraisal value and subjective appraisal result of the improved algorithm are better than that of the conventional non-local average filtering algorithm.  

Fig. (Left to right, up to bottom) image ‘lax’ original image, noisy image, denoised by NLM, denoised by our method noise std=18