Improved regularization nonlocal means model and its application in video denoising

Abstract

Video denoising is the pre-process of most video processing processes. It helps the application having a good data to calculate the accurate solution. Many video denoising methods have been provided in several years in order to create a better input for higher analysis such as segmentation, reconstruction, or super-resolution. The nonlocal means algorithms, the one using the natural redundancy of pattern inside an image to perform a weighted average of pixels whose patches have similar patches to each other is popular nowadays. It is quite adaptable but it still remains two drawbacks: over smoothed the flat areas and leave residual noise at the edges of singular objects. While keeping its benefits and solving its drawbacks, a combination with total minimization has introduced. The model proposed by Rudin, Osher and Fatemi also has its own downsides when they over smoothed the textures, having staircasing effects and contrast losing. Their combination fixes each other problems, and remains their decent traits. The proposed regularizing method is presented by Camille Sutour et at in order to achieve the type of combinations. In the terms of video denoising, they also use 3D Patches provided by Y. Wexler et al. to keep the consistent during the frames transition. Our work is updating the method of 3D patch calculation, so the method can work in varying environment such as rain or snow, or even the losing frames. Keywords: Nonlocal means, Total variation regularization, adaptive filtering, 3D Patches comparing, image and video denoising