Research On Stereoscopic Video Quality Assessment Method For Three-dimensional Television System

Three dimensional (3D) video presents not only content but also depth information of thescene, providing viewers with a more realistic feel of the scene, and3D television (3DTV) whichdelivers3D video is expected as the development direction of next generation of digital television.A complete3DTV system contains the whole chain of3D video capture, representation, coding,transmission, signal conversion as well as display, and it is inevitable to introduce distortions overthe whole processing chain, affecting the perceived quality of3D video. The perceived quality of3D video in3DTV reflects the corresponding video processing technical merits indirectly andperformance of the system directly. Video quality assessment is a key technique in3DTV system;however, research on3D video quality assessment is just beginning. The thesis will study qualityassessment for stereoscopic video (which is currently the most commercialized3D video format),which is mainly from the following three aspects:(1) As traditional video quality metrics are not suitable for evaluating stereoscopic video, wepropose a stereoscopic video quality metric based on traditional metrics. Firstly, factors affectingthe quality of stereoscopic video/image are simplified as left-view video quality, right-view videoquality, and difference between two views from stereoscopic display and viewers’ point of view;then, each of these factors is calculated respectively with traditional metrics for reference; finally,relationships between the factors and the overall perceived quality of stereoscopic video are learnedwith support vector regression (SVR) to predict the objective quality of stereoscopic video.Person’s linear correlation coefficient is higher than0.82for all types of distortions and0.81forindividual distortion in the performance testing, which outperforms the other metrics beencompared and shows that the proposed metric can predict human vision perception for stereoscopicvideo well.(2) As the existing objective quality assessment metrics (which consider the quality ofstereoscopic image as image quality and depth perception) for stereoscopic image need beingweighted for many times, we propose an objective quality assessment metric based on threedimensional-discrete cosine transform (3D-DCT). In the proposed metric, factors affecting thequality of stereoscopic image are combined to a3D signal being processed as a whole, which canavoid unreasonable weighting problems between image quality and depth perception, left-viewimage quality and right-view image quality. When tested in NBU’s and LIVE’s stereoscopic image database, Person’s linear correlation coefficient can achieve0.9314and0.9232respectively, whichoutperforms the six metrics been compared and shows that the proposed metric can predict qualityof stereoscopic image well.(3) As the research on stereoscopic video quality assessment is just beginning, we propose anobjective quality assessment metric for full-HD compressed stereoscopic videos based on theabove quality assessment metric of stereoscopic image. We focus on the characteristics of highdefinition and compression artifacts, propose a perceptual weighting mask to improve the abovemetric of stereoscopic image, and combine the recency effect to the final objective qualityassessment metric for stereoscopic video. When tested with a series of compressed stereoscopicvideos containing H.264artifacts and JPEG2000artifacts, the Person’s linear correlationcoefficient can achieve0.9105, which shows that the proposed metric can predict quality ofcompressed stereoscopic videos well.