Global Opitmization of Vector Quantizer Based Communication Systems

9314335玫瑰这个项目解决了音频和视频信号的数字通信矢量编码的基本问题。整个信号压缩问题被重新表述为一个广义矢量量化器的设计问题，其中包括信号分解，量化和重建。攻击的主线，使用这种新的范例，是嵌入的概率框架内的确定性退火，这是基于最大熵原理及其推广，交叉熵原理的基础上得出的优化问题。这产生了两个主要好处：避免陷入贪婪方法的局部最优的能力，以及通过引入额外约束对各种目标的联合优化的可扩展性。该项目的第一阶段将包括优化对数据通信特别重要的孤立问题：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增长。第一阶段预计将产生实质性的改进，目前使用的方法，并在同一时间提供的基础上，第二阶段，其中包括合成不同的优化问题。该项目的第二阶段将包括联合和同时优化第一阶段解决的问题的子集，从而消除由于独立分段优化的次优。要解决的一个特别重要的特殊情况是针对给定源和有噪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为在严格的B和有限的信道（如未来一代蜂窝网络）的苛刻环境中实现可靠和高效的通信提供理论上令人兴奋但实际上有用的解决方案。为了验证本项目开发的新优化方法，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。将确定性能改进的基准，目的是使实际信号压缩系统的未来设计者相信新技术的实际效用。特别是，两个广泛研究的问题，量化的LPC参数的频谱建模语音编码和低速率变换编码的图像作为基准性能的目标。在一种或两种情况下，信道误差的模型将被包括在内，以评估整体信源信道编码性能。***

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. ***