Prioritized video coding for ATM networks

The interface between variable bit rate video coders and asynchronous-transfer-mode (ATM) networks has significant bearing upon network performance and quality-of-service (QoS). This is particularly true during network congestion where bandwidth resources are scarce. One network proposal for mitigating the effects of congestion is prioritization. Under prioritization, cells are dropped or delayed according to the importance of the data they carry. This thesis focuses upon the relationship between the prioritized video coding algorithm and the network traffic characteristics to provide graceful degradation of service during network congestion. We begin by mathematically formalizing prioritization within the framework of probability and information theory. This analysis reveals the effects of prioritization upon the moments and distribution of ATM cells. In addition, a general algorithm is derived for estimating the bit rate generated at a particular QoS which is ideally suited for non-stationary, multidimensional data such as video and is computationally efficient. Having identified the role coder prioritization plays upon network performance, we focus on the various forms of prioritization. We begin by comparing a number of block-based prioritization techniques such that the high priority data stream is MPEG-I compliant. These techniques are compared in terms of their cell distribution, temporal characteristics, computational efficiency, and subjective quality as well as the mathematical predictions performed previously. By relaxing the MPEG compliance constraint, the thesis then investigates multiresolution algorithms for analyzing the video prior to prioritization. We begin by demonstrating the value of subband coding and propose a modified MPEG coding algorithm and demonstrate its superior performance. Based on this experience, we propose a new class of linear transformations called multiscale transforms which combine the performance characteristics of block-based coding with a notion of scale. A pair of design algorithms are presented based on a unique decomposition of the autocorrelation matrix in terms of its subsampled cross-correlations. The superior performance of the multiscale transform over the block DCT is illustrated for a first-order Markov process as well as a prioritized video