Image Denoising Feedback Filter Based Residuals TV Model

Vision is the most important source of information, images are almost everywhere. But images will inevitably be polluted by the noise in the process of acquisition, transmission and display. Image de-noising has been the most fundamental, core problem in the field of image processing. In many application areas, the image detail information such as edge and texture is very important. So the research of image de-noising methods, which can keep not only the image texture details, but also the image smooth area good visual effect, is the current hot. ROF model of TV is a classical model for image de-noising, it can keep the image edge and the smooth area good visual effect, but at the same time filter a lot of texture details. Some methods of keeping texture details are bad for keeping edge, or have big amount of calculation.Because of TV filtering a lot of texture details, we extract the texture from the residual image on the basis of TV model, in order to obtain image edge, while retaining the texture, and reducing the algorithm complexity. The thought proposed in this paper:firstly we use TV model to divide the image into cartoon part and residual part, then we extract texture from the residual part. Finally add the texture details back to cartoon part. We put forward two methods:the first is to add texture details which extracted by DWT and DCT directly back to the cartoon part; the second method is that, in the framework of Bregman iteration we add the extracted texture details back to the residual image in every step of the iterative process. In the first method, we propose different thresholds when extracting the texture in the residual image by using DWT and DCT. From the de-noising images which are extracted from residual image by DCT, we can see that DCT is more effective than TV both in terms of visual effect and in the numerical of PSNR and SSIM. Then we propose some modification by local smoothing about the block effects by using DCT. At last, in the framework of Bregman iteration, we add back the residual image after using DCT extraction in the process of iteration. Experiments show that compared to Bregman iteration this method proposed in this paper make not only a smooth area good visual effect, but higher PSNR and SSIM values, and better iterative stability.