改进对向传播神经网络应用于码书设计及其性能分析

Modified CounterPropagation Network Applied to Codebook Design and Its Performance Analysis

针对对向传播神经网络(CPN)应用于矢量量化时的两个缺陷进行改进,提出了一种码书设计算法——快速竞争学习及误差修正算法(FCLECA),并设计了相应的基于改进CPN的快速矢量量化器模型,详细讨论了FCLECA的重要步骤、重要参数及其时间复杂度仿真实验结果表明文中算法能在提高码书质量的同时大幅缩短训练时间。

A new fast codebook designing algorithm referred to as FCLECA (fast competitive learning and error correction algorithm) and based on a modified CPN(counterpropagation network) is presented. The corresponding fast vector quantizer model based on the modified CPN is illustrated. Details of the critical steps and parameters in FCLECA and its time complexity are discussed. Results of simulating experiments indicate that with the new algorithm the training time in codebook designing is greatly reduced, while quality of the generated codebooks is improved. The new algorithm is effective and feasible in vector quantization.