Global Opitmization of Vector Quantizer Based Communication Systems

9314335 Rose这个项目解决了用于音频和视觉信号数字通信的矢量编码的基本问题。整体的信号压缩问题被重新表述为一个广义的矢量量化器设计问题，其中包括信号分解，量化和重建。使用这种新范式的主要攻击路线是将优化问题嵌入确定性退火的概率框架中，确定性退火是基于最大熵原理及其推广，即交叉熵原理推导出来的。这产生了两个主要的好处：能够避免陷入贪心方法的一些局部最优，以及通过合并额外的约束来扩展到各种目标的联合优化。该项目的第一阶段将包括对数据通信特别重要的孤立问题的优化：源信道矢量编码、熵约束矢量量化、结构约束矢量量化、广义矢量量化和不平衡树矢量量化器生长。第一阶段预计将对目前使用的方法产生实质性的改进，同时为第二阶段提供基础，第二阶段包括综合不同的优化问题。该项目的第二阶段将包括对第一阶段所解决问题的子集进行联合和同时优化，从而消除由于独立分段优化而导致的次优性。要解决的一个特别重要的特殊情况是给定源和噪声信道的矢量量化器设计的组合优化问题，受编码复杂性和比特率约束。一种以非启发式方式同时解决这些问题的方法有望提供一个理论上令人兴奋但实际上有用的解决方案，以在苛刻的和有限的信道环境中实现可靠和有效的通信，例如下一代蜂窝网络。为了验证本项目中开发的新优化方法，将对使用这些方法的特定语音和图像压缩目标进行测试和模拟。性能改进的基准将被确定，目的是说服未来的实际信号压缩系统的设计者的新技术的实际效用。特别是两个被广泛研究的问题，语音编码中用于频谱建模的LPC参数的量化和图像的低速率变换编码，作为基准性能的目标。在一种或两种情况下，将包括信道错误模型来评估总体源信道编码性能。* * *

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. ***