提高PCA识别率的新算法

A new PCA method to increase recognition rate

针对基于主元分析(PCA)的识别算法不能最优区分不同种类样本的缺点,提出了一种新的多主元分析(Multi-PCA)识别算法。该算法为每类样本构造单独特征空间,用各个空间的特征向量重建待识别样本。特征空间是基于某类样本图像的共性建立,因此重建该类样本图像时将得到较小重建误差,而重建其它类图像时的误差较大。可以根据重建误差的大小来识别样本图像,将待识别样本分类到具有最小重建误差的特征空间。在ORL、YALE人脸库和标牌压印字符库上的实验结果显示,Multi-PCA的识别率比PCA有较大提高。

The PCA（Principal component analysis）is not the best method to extract features for recognition because the difference between different kinds is not considered.So a new Multi-PCA method is proposed for recognition application. The feature spaces are established of every kind sample based on the commonness of a specifically kind of samples.Then a new sample to be recognized is reconstructed using features vectors of every feature space and the reconstructed error is calculated.If a sample belongs to a specifically kind then the reconstructed error of this sample will be small and the error will be enlarged if it belongs to another kind.Based on this principle the samples are classified to the space in which the error is the least.Experiments on ORL and YALE faces databases and scutcheon character database show that the method is more accurate than the standard PCA.