Image Denoising Based On Regional Division

Image denoising has always been one of the research focuses in the field of computer image processing and computer vision. As the developments of imaging analysis and fellow-up applications on images, the requirements for image quality have been increasingly high, which leads to the uninterrupted researches on image denoising since it appeared. But the algorithms proposed before all have some problems themselves, even though they do have their denoising effects. Recently, non-local means denoising method proposed by Antonu Buades, and BM3D denoising algorithm proposed by Dabov both fully make use of the similarity of images and get good denoising effects. Howerver, artificial traces in the smooth are still obviously.Though BM3D method has been considered as the best denoising method, it has its weak point as the plaque effect in the smooth regions. In order to combine the advantages of algorithms, image denoising based on regional division is proposed in this paper. First, a noisy image is divided into many small blocks. It is easily found that the variance of the blocks are different between the smooth regions and non-smooth regions of the noisy image by lots of experiments. In other words, the variance in smooth regions is small, but the variance in non-smooth regions is very large. The blocks in which the variance is less than a certain threshold are considered as smooth regions, and the remaining blocks are consider as non-smooth regions. In this way, a noisy image is divided into smooth regions and non-smooth regions.Second, primal sketch is a sparse representation model, which divide an image into two components,"sketchable" and "non-sketchable" parts. The sketchable part just contains the strture of the image. The direction of every pixel in primal sketch is used for structure block to get the strture of the image.Finally, an image can divide into smooth and non-smooth regions by variance statistice, and structural and non-structural parts by primal sketch sparse representation model. The structural part may overlap with the smooth regions. So we make a fusion rule that if a block overlaps with the structural part, it is processed by structural part’s method, otherwise by the original smooth region’s method and non-smooth region’s method. Thus the image is divided into structural areas, the smooth regions and the non-smooth regions.For the smooth region, as non-local mean method leaves artificial traces in the smooth region, that is the phenomenon of "Pseudo-texture". The non-local mean method is added with the mean ideological in this paper so as to eliminate the phenomenon of "Pseudo-texture", which produces better effects. For the structural region, the way in which the blocks are taken in the BM3D method is improved. The taken blocks are no longer the traditional square blocks but the square blocks whose orients are consistent with the Primal Sketch segments. These blocks are combined into3-dimentional data, and denoising is then carried on by3-dimentional filtering method.We analysis the difference of the original methods and the improved method in denoising. And experiments prove the effectiveness and feasibility of the proposed method.Finally, we summary the main works in this paper and give the future works are given.