Quantitative Analysis of GoP Structure for Scalable Video Coding: an Information Theory Perspective

Abstract

As the demand for real-time video streaming fast rising, the scalable video coding (SVC) is widely recognized in both academic and industrial societies. In SVC, group-of-picture (GoP) structure is a base unit consisting of a sequence of frames and its structure contains a periodic relationship among them. Hence, the GoP structure not only determines the efficiency of SVC encoder, but also impacts on the video quality of a decoder output. However, the conventional representation systems could not clearly represent the GoP structure for both the encoder and decoder. Additionally, there is lack of theoretical study on the achievable performance of the encoder/decoder. To address this, this paper focuses on quantitative analysis of the GoP structure from the perspective of information theory. We propose to represent the GoP structure by introducing and extending the probabilistic graphic model which is proposed as a methodology in machine learning society. By the proposed representation system, the information compression of video is formulated for the encoder/decoder with closed-form expression. Based on this, we quantify the video information for GoP and propose a methodology to optimize the achievable performance of a given GoP in the view of information theory. The proposed representation and optimization methods are valuable to design an efficient GoP structure.