利用遗传算法实现手写体数字识别中特征维数的压缩

SELECTING FEATURES WITH GENETIC ALGORITHM IN HANDPRINT DIGITS RECOGNITION

本文提出一种模式识别中压缩特征维数的方法,首先从信息论观点出发,以类-特征和特征-特征之间的互信息构造综合互信息指标,选择类-特征互信息高而特征-特征互信息低的特征子集实现维数的压缩.进而利用起源于生物进化论观点的遗传算法辅助进行这样的特征选择.设计了染色体的表示方式、适应度公式和两种新的遗传操作算子.通过对手写体数字的识别实验表明,这种方法在不显著降低识别效果的前提下有效地减少了特征的维数,简化了分类器的设计.

A feature dimension compression method is described. In a real pattern recognition system, a large number of features usually make the realization of efficient classifier a difficulty. The redundancy within feature set is also unavoidable. The method proposed in this paper selects features from the existing feature set according to the mutual information measurement between classes and features. Genetic Algorithm(GA) is used to select the most informative feature subset. GA is a kind of optimization algorithms originated from the Theory of Natural Evolution. Based on the experiment results of handprint digits recognition, the proposed method can reduce the mumber of features to be used in the recognition process and without impairing the correction rates significantly.