Blind configuration of multi-view video coder prediction structure

摘要

Efficient coding of 3D multi-view video depends on the group of pictures (GOP) prediction structure and the video stream encoding order. Optimizing the GOP prediction structure and the stream coding order will reduce the coding bit rate, improve the peak signal to noise ratio (PSNR) and reduce the coding complexity. To date, conventional coders are manually configured based on prior knowledge of the geometric arrangement of the video cameras and the properties of the video streams. In this paper, a blind self-configurable multi-view video coder (BC-MVC) algorithm is introduced. The proposed BC-MVC blindly estimates a GOP prediction structure without prior knowledge of the cameras' geometric arrangement. The BC-MVC decomposes the key video frames into independent bases and a projection (mixing) matrix using blind source separation. Based on the mixing matrix, an algorithm is developed to estimate the cameras' geometric arrangement and consequently an optimum GOP prediction structure. The experimental results show that the proposed blind multi-view video coder has better coding efficiency than conventional 3D multi-view video coders with predefined coding structures. It also shows that BC-MVC is robust to camera failures and severe channel errors. Moreover, the numerical complexity analysis shows that the proposed BC-MVC algorithm has lower computational complexity than existing multi-view video prediction schemes.1