依概率分类的保持投影及其在人脸识别中的应用

Probabilistic classification preserving projections and its application to face recognition

为了提高低维空间对原始高维样本的表示能力,该文提出了依概率分类的保持投影算法(PCPP)。PCPP考虑了样本类别信息,并重新定义类内样本间的相似性,包含样本的邻域信息,而且在K近邻选择下,还能反映样本被正确归类的概率。样本经投影后,在低维特征空间内,被正确归类且概率较大的类内样本间的邻域关系得到了保持。在Yale、FERET及AR人脸库上的人脸识别实验表明,PCPP较其他算法取得了更好的识别性能。

In order to improve the ability of low dimensional space to represent high-dimensional samples,a novel manifold learning method called probabilistic classification maintain projection(PCPP) is proposed.The PCPP takes class information into account and refines similarity weights of intra-class samples,which not only contain neighborhood information of samples,but also can reflect the probability that a sample can be correctly classified when its K nearest neighbors are selected.After projection,neighborhood relationship of the intra-class samples which possess more classification probability can be preserved.Experimental results on the Yale,FERET and AR face databases demonstrate that the PCPP performs better than other algorithms.