An Image Denoising Method Based On Primal Sketch Amendment And Low-rank Model

Image noise reduction as a research focus of computer image processing andcomputer vision fields, has been driving researchers to search for more effective imagenoise reduction methods. Currently, there are many outstanding performace noisereduction methods, such as Non-Local Means, K-SVD, BM3D, but these methods donot distinguish the different areas of an natural image (usually including smooth region,the non-smooth region, structural areas, etc.), just treat them with the same processingflow, resulting in the appearance of such phenomenons: the smooth areas of an imagelook obvious artifactly or with a high calculation complexity.In order to improve above mentioned phenomenons, this paper proposes a noisereduction method based on Primal Sketch amendment and low-rank model, whichbasic idea is that: for a noisy image, when its noise intensity value is not greater than30, directly imposes Primal Sketch to get its initial sketches; and when the value isgreater than30, firstly uses the previous image noise reduction to process imagereconstruction based on wavelet transform and then imposes Primal Sketch to get itsinitial sketch. A two-stage sketch line amendment rules for the region structure isapplied to amend the initial sketch aboved, for the amended sketch’s segments on eachpoint along its line to design direction of the sketch window in the direction of the lineand get its region image. The region image is mapped to the structure in accordancewith regional map regional and non-structural areas, by using the image block variancestatistical method the image is divided into smooth and non-smooth region area. Theoverlapping regions between smooth regions and non-smooth、non-structural regionswas divided as the ultimate smooth regions and non-smooth region, and keep theoriginal structure of the regional division as the final division result, thus the noisyimage is divided into structural area, regional and non-smooth smooth area. Forsmooth regions using non-local means method based on average block, and forstructural regions, non-smooth regions, are based on the use of the present inventionprovides a noise reduction method based on matrix completion, core idea of thismethod is that the problem of image denoising was converted to the problem of matrixcompletion. Finally, combines the denoising results of different regions to obtain thefinal denoised image.By comparing and analysing the differences among this method, the noisereduction method based on matrix completion and several classic noise reduction methods in noise reduction performance, we demonstrate the effectiveness andfeasibility of the proposed image noise reduction method, which not only effectivelyimprove the division accuracy of image regions, but also with a more substantialefficiency increment, makes the final noise reduction result better than the noisereduction method based on regional devision and common classical noise reductionmethods.