最大熵矢量量化及其在TMS320DM642上的实现

Maximum Entropy Vector Quantization and Its Implementation on TMS320DM642

为了克服传统的矢量量化方法存在的信息损失量大及二次规划(Quadratic pragramming,QP)量化方法计算复杂度大等缺点,提出了一种新的量化方法——最大熵量化。这种量化方法一方面能将量化权值的熵最大化,从而确保在没有先验知识的情况下不会造成太多量化误差;另一方面则考虑了矢量集合在时间空间上的分布关系。本文在TMS320DM642处理器上实现了这种算法,并进行了一系列的算法和程序层的优化。在基于图像的目标识别应用中的实验证明,最大熵矢量量化算法及其在TMS320DM642上的实现,不仅能提高识别的性能,而且能满足实时性的要求。

To overcome the shortages of traditional vector quantization（VQ） and quadratic pro- gramming （QP） VQ algorithm, a new quantitative method i.e. the maximum entropy quantiza- tion is proposed. The quantization approach can maximize the entropy of the quantification weights, thus avoiding too much quantization error when a priori knowledge is absent. And it considers the space-time distribution of vector collection. The TMS320DM642 processors are used to implement the algorithm with a series of optimization on algorithm and program levels. Experiments on object recognition prove that the maximum entropy vector quantization algo- rithms with acceleration on TMS320DM642 processor can improve the recognition performance and meets the real-time requirements.