Beyond Traditional Transform Compression: Universality and Multiple Description Coding
Thursday, March 12, 1998

Transform coding, including both block discrete cosine transform (DCT) and wavelet transform methods, is the basis for the JPEG and MPEG standards for image and video coding. On the other hand, information theoretic bounds on source coding use, implicitly or explicitly, vector quantization. What then is the reason for the dominance of transform coding in applications? We extend transform coding to the problems of universal lossy coding and coding for erasure channels since transform coding has lower computational complexity for a given vector dimension, allowing the practical use of larger vectors. In universal coding, one wishes to design an algorithm which performs well for all the sources in some class of sources. I will present two algorithms for coding Gaussian sources with unknown correlation. Subject to certain conditions, these algorithms give performance which approach the performance of an optimal transform coder designed with knowledge of the source correlation. Modern packet communications systems use CRC bits for error detection and discard packets with bit errors. Since packets are either received correctly or lost completely, these networks can be effectively modeled as packet erasure channels. For general data communication, erasures are handled by retransmission. However, retransmission is not feasible for broadcast; when feedback is unavailable or expensive; or when there are stringent delay constraints, such as for streaming audio and video. Coding for erasure channels without retransmission can be seen as a generalization of the multiple description problem in information theory. I will present two coding methods for erasure channels based on transform coding. These methods, which combine source and channel coding, have the desirable property that as the number of packets lost increases, the performance degrades gracefully. In this respect they are superior to separate source and channel coding.
UC Berkeley Networking
Ashwin Pananjady and Orhan Ocal
Last Modification Date: Wednesday, February 10, 2016