Video De-noising using graph signal processing

Abstract

"Depth map videos have distinct characteristics and usually they are known as piecewise smooth (PWS) videos. Due to this characteristics, they are different from the natural videos as they have objects with sharp edges and no texture. Most of the denoising techniques have been designed for natural images and videos. That is why they are not suitable for depth map videos. The Graph signal processing is emerging field and recently there is work done in the area of image processing, for example, Non-local Graph based transform (NLGBT). But there is no work done in the area of video processing. We proposed an extension of NLGBT for videos in this thesis and called it as Video NLGBT. One of the important application of depth map videos is Depth Image Based Rendering (DIBR) in 3DTV. In DIBR based 3DTV, viewers will not see depth map videos but they will see synthesized views generated using DIBR. So doing the quality assessment on denoised depth map videos is not a good idea. Ultimately what viewers experience in 3DTV is important. So we did a quality assessment on synthesized videos in this thesis. As DIBR introduce geometric transformation of edges and objects, traditional quality assessment metrics may not be useful because they are doing comparison pixel by pixel. So we used MW-PSNR for quality assessment which accommodates geometric transformation of DIBR synthesized videos. In this thesis we did a quality assessment of Video NLGBT and compared it with state of the art algorithm V-BM3D."