Global Opitmization of Vector Quantizer Based Communication Systems

点击翻译按钮获取中文摘要

9314335 Rose This project addresses fundamental problems in vector coding for digital communication of audio and visual signals. The overall signal compression problem is reformulated as a generalized vector quantizer design problem, which incorporates signal decomposition, quantization, and reconstruction. The main line of attack, using this new paradigm, is to embed the optimization problem within the probabilistic framework of deterministic annealing, which is derived based on the principle of maximum entropy and its generalization, the principle of cross-entropy. This yields two main benefits: ability to avoid some of the local optima that trap greedy methods, and extendibility to joint optimization of various objectives by incorporating additional constraints. The first phase of the project will include optimization of isolated problems of particular importance to data communications: source-channel vector coding, entropy constrained vector quantization, structurally constrained vector quantization, generalized vector quantization, and unbalanced tree vector quantizer growing. The first phase is expected to yield substantial improvements over methods currently in use, and at the same time provide the basis for the second phase which consists of synthesizing the different optimization problems. The second phase of the project will include joint and simultaneous optimization of subsets of the problems addressed in the first phase, thus removing the suboptimality due to independent piecewise optimization. A particularly important special case to be addressed is the combined optimization problem of a vector quantizer design for a given source and noisy channel, subject to encoding complexity and bit rate constraints. A method addressing these issues simultaneously in a nonheuristic manner is expected to provide a theoretically exciting yet practically useful solution to reliable and efficient communications in the demanding context of critically b andlimited channels, such as the future generation of cellular networks. In order to validate the new optimization methods developed in this project, specific speech and image compression targets utilizing these methods will be tested and simulate. Benchmarks of performance improvement will be determined with the intent of convincing future designers of practical signal compression systems of the practical utility of the new techniques. In particular, two widely studied problems, the quantization of LPC parameters for spectral modeling in speech coding and low-rate transform coding of images with serve as targets for benchmarking performance. In one or both cases, a model of channel errors will be included to assess overall source-channel coding performance. ***