代价敏感学习的稀疏局部保留投影算法  

Cost-sensitive Sparse Locality Preserving Projections

作　　者：林克正[1] 钟岩[1] 程卫月[1] 

机构地区：[1]哈尔滨理工大学计算机科学与技术学院,黑龙江哈尔滨150080

出　　处：《哈尔滨理工大学学报》2015年第3期45-50,共6页Journal of Harbin University of Science and Technology

基　　金：黑龙江省教育厅科学技术研究项目(11551087)

摘　　要：针对已有的局部保留投影(locality preserving projections,LPP)算法可能将相似的类别误投影到一起,导致正确识别率降低的问题.在局部保留投影算法的基础上,提出了一种基于代价敏感学习的稀疏局部保留投影算法(cost-sensitive sparse locality preserving projections,CSLPP).该算法将代价敏感学习引入到人脸识别中,首先对样本进行代价敏感思考,然后再将样本稀疏化,最后求得最优投影向量.通过在YALE人脸库和FERET人脸库上实验,结果表明CSLPP算法在投影之前将代价考虑进去,有效的避免了高风险,该算法在最近邻分类器上的的识别率明显高出其它算法的识别率.In Locality Preserving Projections algorithm, faces in similar categories are projected as the same one, leading to the decrease of recognition rate. To solve this problem, Cost-sensitive Sparse Locality Preserving Projections algorithmbased on LPP algorithm is proposed. In CSLPP algorithm, in which Cost-Sensitive Learning was applied to face recognition, face samples were first cost-sensitively thought of, and Sparseness, at last the opti- mal projection vector was figured out. Experimental results on the YALE and FERET face databases show that CS- LPP algorithm effectively avoids high risks and its recognition rate is significantly higher than that of others in Nea- rest Neighbor Classifier.

关 键 词：局部保留投影 人脸识别 代价敏感 稀疏化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]