A Study On Perceptual Stereoscopic Video Coding Based On Disparity-based Just-noticeable-distortion Models

In recent years, three-dimensional television(3DTV) technologies have advanced rapidly due to the demands of the realistic visual experiences. Multi-view videos which can give users more vivid experiences are generated by capturing the same scene with several cameras simultaneously at different locations. Nevertheless, with the growth of cameras number, the storage space and bandwidth will increase exponentially. It is not conducive to store and transmit stereoscope videos. Therefore, effective stereo video coding is needed. Because the final receiver of the video signals is human visual system generally, the combinations of visual perception factors and video coding have draw lots of attention. In this thesis, the concept of JND models and how to use them into image or video coding are discussed based upon the H.264 and multi-view video coding(MVC)theory. The strengths and weaknesses have been analyzed respectively in the thesis.Disparity based JND models are proposed by considering the relationship of disparity and human visual sensitivity. One of them considers the stereo matching in order to get more accurate disparity map. And a chrominance domain JND model is proposed. Then,the other one combines the depth of focus(DOF) blur effect with human visual features to improve the JND model. The proposed model can not only saving bit-rate but also improving the image or video perception quality. Our thesis has launched discussions on the proposed JND models and the combination of them with MVC. We undertake our mainly tasks over following areas:1. Disparity-based JND model is presented in chapter 4. Image segmentation based stereo matching can get more accurate disparity information. The traditional spacial and temporal JND model are improved by considering the relationship of disparity and human visual sensitivity. In order to prove the accuracy of the JND model, we use it on MVC. The experiment results show this method effectively removing the disparity redundancy and saving bit-rate.2. Chrominance JND model is put forward by taking into the consideration of human color perception feature. It is estimated by Gaussian distribution which can simulate the distribution of cones on retina. The experiment prove this method significantly save bit-rate and remove color redundancy without decreasing of color perception.3. Depth of focus blur(DOF) based JND model is proposed. It combines the DOF blur effect with JND model. Using the disparity information to divide the foreground and background regions. For the different regions, using the JND model to choose the different quantization parameters. It can effectively share out the bit-rate of the background regions to front regions. Experiments demonstrate that the proposed method effectively improve the visual comfort in stereoscopic viewing while maintaining the bit-rate without increasing.