Research On Depth Extraction Of Natural Three-dimensional Television System

With the development of television(TV) which from black and white TV, color TV to digital high-definition TV, the TV technology has gradually towards the direction of large-screen, color, high-definition and multimedia. The three-dimensional television (3DTV) has a depth and ambiance sense, so it will surely to be the next direction of TV. It has a great practical significance to research the 3DTV system.Around the status of 3DTV research, the paper propose a new 3DTV system - a natural three-dimensional television system. The system is stressed that the natural scene and natural 3D feeling without wearing glasses, and using the currently practical coding structure of "one video + depth" approach, while introduced the concept of "occlusion information". With the "occlusion information", not only do the audiences have the clear 3D display of main view, but also providing a rich side and occlusion information. Additional, this system has good scalability, which can obtain the different quality of 3D effect by configuring the camera, blocking information and control bandwidth and storage costs.Then, I have done a lot of work in the key challenges of depth extracting of the natural 3DTV system. After researching the status and theory of stereo matching algorithm, I programmed a experimental platform software for stereo matching algorithms and achieved a wide range of stereo matching algorithm.Finally, Inspired by the defects of single stereo matching algorithm, I propose a novel parallel optimization model for two arbitrary stereo matching algorithms. It includes four steps: Firstly, reliable disparity pixel marking; Secondly, reliable disparity layer selecting; And then, unreliable region filling based on cross aggregation; Finally, border handling based on canny edge detection. The experimental results show that the model can make full use of the advantages of two methods, especially for both poor methods. Additionally, our model is very efficient and it takes less than one second for a typical image. Therefore, it can be used as an optimization model for both real-time stereo parallel algorithms.