Video compression with complete information for pre-recorded sources

Traditional video compression algorithms focus on causal processing of the video data. The causal constraint is necessary for real-time video programs where delay is an important consideration. However, many video programs such as movies are pre-recorded and can be processed offline prior to compression. This thesis proposes using pre-processing for pre-recorded video programs to obtain useful information for compression. By exploiting information obtained about a program prior to compression, better video quality can be achieved. In particular, as an example of how complete noncausal knowledge of a video program can be used to improve performance over causal knowledge, a new iterative algorithm is developed for a buffer-constrained quantization problem to reduce the number of quantization changes in MPEG-2 intraframe video compression. This algorithm is shown to be optimal under certain conditions. By introducing a noncausally computed parameter into the distortion function, experiments have shown that improved and more constant video quality can be delivered using the noncausal approach relative to a traditional causal approach. Gains of up to 1.0 dB in PSNR were observed in multiscene sequences, corresponding to savings of up to 10% in bit rate. Thesis Supervisor: Jae S. Lim Title: Professor of Electrical Engineering