Image Denoising Method Based On Gradient Histogram Preservation Model

In the computer vision point of view, the image noise can significantly reducethe visual effect of the image, and affect the performance of subsequent processingand analysis algorithms. Therefore, image denoising as a kind of typical underlyingvision problem has attracted wide attention of many scholars. For this reason, peoplehave proposed a lot of image denoising methods, such as total variation (TV)method, the non-local centralized sparse representation (NCSR) method, threedimensional block matching (BM3D) method. However, among all of thesedenoising methods, fine or small-scale textures of the original image can not befully preserved in the denoising process. On the basis of how to keep these finetextures or small scale structures, removing image noise effectively is still one of theimportant issues to be solved in this area.Now many priors have been widely used in the image denoising methods. Onerepresentative class of the image priors is the gradient priors. Research of naturalimages shows that the image gradient obeys heavy-tailed distribution. In this paper,we use the hyper-Laplacian distribution to fit this heavy-tailed distribution of thenatural images. However, these prior-based denoising methods tend to smooth outthe fine scale texture structures while removing noise, thereby reducing the visualquality of the image. To address this problem, in this paper we propose a textureenhanced image denoising method that uses a gradient histogram preservation model,enforcing the gradient distribution of the denoised image to be as much as possibleclose to the estimated gradient distribution from the original image.On the other hand, different image priors describe the different statisticalproperties, therefore it is possible to combine different image priors together toimprove the image denoising effect. In this paper, the image denoising method is acombination of a hyper-Laplacian prior and gradient histogram preservation model,using the iterative histogram specification algorithm for solving the problem.Experimental results show that our method gets a greater PSNR value and SSIMvalue, and can well keep the detailed texture structures of the denoised image,making the denoised image look more natural.An image contains a lot of different areas of different textures. Their gradientdistributions have obvious differences, interfering with each other. For example, anarea rich in texture information will affect the denoising effect of the area withrelatively simple texture information, resulting in the wrong textures. To solve thisproblem, we propose a denoising method based on region segmentation. Compared with the image denoising method without segmentaton, we can see that, the imagedenoising method based on regioin segmentation can preserve the fine textures andget higher visual quality.The above gradient histogram preservation model can also be extended tosolve the image deblurring problem, using alternating minimization method.Deblurring results show that our proposed method preseves the delicate textures ofthe images, getting sharper images and better visual effects.