论文部分内容阅读
提出了误差竞争学习 ( Distortion competitive learning,DCL)算法。该算法基于 Gersho的矢量量化误差渐近理论的等误差原则 ,即当码本数趋于无穷大时 ,各区域子误差相等 ,使用这个原则作为最优码书设计的一个必要条件 ,并结合传统最优码书设计的两个必要条件 ,然后根据这 3个必要条件 :( 1)最近邻规则 ;( 2 )中心准则 ;( 3)各区域子误差近似相等设计最优码书 ,而在算法的实现中引入广义误差测度 ,以确保该测度与各个区域的子误差相关。最后从快速性与均方差两个方面与目前的码本设计算法比较。实验结果表明 ,该算法在降低迭代时间与减少均方差优于其他传统码本设计算法 ,能迅速地找到优化的码本。
The algorithm of Distortion competitive learning (DCL) is proposed. The algorithm is based on the principle of equal error of Gersho’s vector quantization error asymptotic theory. That is, when the codebook number approaches infinity, the sub-errors of all regions are equal. Using this principle as a necessary condition for the optimal codebook design, Then, according to these three necessary conditions: (1) nearest neighbor rule; (2) center criterion; (3) sub-error of each region is approximately equal to the optimal codebook, and in the realization of the algorithm Generalized error measure is introduced to ensure that the measure is related to the sub-error in each area. Finally, compared with the current codebook design algorithm from two aspects of fastness and mean square error. The experimental results show that the proposed algorithm can reduce the iterative time and reduce the mean square error better than other traditional codebook design algorithms, and can quickly find the optimized codebook.