Research On The Key Technologies Of Stereo Video Quality Assessment

With the broadcasting of the3DTV channel in domestic and foreign, stereo videohas been gradually into the ordinary families. The video which can provide peoplewith stereoscopic perceptions includes two types. One is the two-view stereo videowhich is currently used by the3D channel, in which the viewers need to wear special3D glasses at the receiver. The other is called multiview stereo video which iscomposed of more than two viewpoints that enable3D viewing for multiple userswithout requiring3D glasses at autostereoscopic displays. To enable3DTV and FTVin real-world applied, the entire processing chain including3D contents capturing,preprocessing, coding, transmission, decoding/rendering, and displaying should beconsidered. As a consequence a variety of distortions may occur in this chain whichwill impair the quality of the3D scene, especially the compression coding stage. Soassessing the quality of the stereo video is very important. However, the stereo videois different from the general2D video, ordinary2D evaluation metrics can not bedirectly used for3D video. How to assess the stereo video quality has become a veryimportant and urgent issue to its development. Existing objective stereo videoquality assessment methods are better than subjective metric which needs specialenvironment, time-consuming, and not real-time. However, the fitting with thesubjective evaluation results is poor, and can not well reflect the human visualcharacteristics. In this paper, the metrics and key technologies of stereo videoobjective quality assessment are studied. The major contributions of this paper are asfollows:1) Traditional JND model is improved according to the characteristics of stereovideos to establish a BJND model, and a stereo video quality assessment metricnamed BPSPNR are proposed. Compared with PSNR and PSPNR, the results ofproposed method are closer to the subjective test judgments whether from the correlation or the dispersion of the data sample.2) Three Full-Reference quality assessment metrics are proposed.(1) For themultiview video, an objective assessment model combining2D assessing metricswith the view difference maps is proposed to analyze the performance of PSNR,SSIM, and VQM. Experimental results prove that SSIM is most suitable for thevideo quality evaluation, PSNR is most suitable for the perceived stereo quality,while VQM is not suitable for the stereo perception evaluation.(2) For the stereoimage fused by two views, aiming at the human visual features of depth perceptionand regions of interest (ROI), a novel assessing method based on the ROI of textureand depth map is proposed. The ROI of texture image and the corresponding depthmap are extracted by using visual attention extraction tool. The weight factors areallocated according to the degree of interesting of each region. This method caneffectively reflect human subjective perception for the stereo image.(3) For thesynthesized virtual view, an assessing metric based on edge difference is proposedBased on the analysis on the pixel difference between virtual and original views,each difference pixel is classified and assigned with a visual weight, and higherweights are applied to the edge pixels. Experiments for multiview video sequencesprove that the results of this metric are more in accordance with the characteristic ofhuman visual system compared to other assessment methods.3) A Non-Reference assessment metric is proposed for the virtual viewsynthesized by multiview video plus depth (MVD) format. The method compares theedge similarity of the depth and the corresponding texture maps which generate theintermediate virtual view, and combines it with the virtual views’ edge blockiness toevaluate the quality of the virtual view. Experiment results show that the proposedmethod can better reflect the quality of the virtual view.4) According to human visual sensitivity characteristic on spatial and temporaldepth changes, we study the just noticeable difference (JNDD) of the depth detailperception. JNDD model is proposed through derivation and analysis on a large number of subjective experiments. Then, the JNDD model is applied to optimize thedepth map. The proposed trilateral filtering algorithm based on JNDD can smooththe object edges and reduce the edge noise in synthesized view while reduce thebitrates, which verifies the rationality of the proposed JNDD.