基于分块聚类的多流形判别分析人脸识别  

MMDA with blocking clustering for face recognition

作　　者：王丽艳[1] 李伟生[2] 

机构地区：[1]重庆邮电大学移通学院计算机科学系,重庆401520 [2]重庆邮电大学计算机科学与技术学院,重庆400065

出　　处：《计算机应用研究》2014年第9期2853-2855,共3页Application Research of Computers

基　　金：国家自然科学基金资助项目(60842003)

摘　　要：针对传统的人脸识别算法在处理单样本人脸识别时由于训练样本不足而影响识别率的问题,提出了一种基于分块聚类的多流形判别分析(MMDA)算法。将每个单训练样本划分成若干大小相等且互不重叠的局部小块,利用聚类算法将局部小块聚类到各个类所属的流形上,并使用特征变换最大化类与类之间的分离性;最后,计算出测试人脸的流形与所有训练样本流形之间的距离,采用最近邻分类器完成人脸的识别。在ORL及FERET两大人脸数据库上的实验验证了算法的有效性及可靠性,识别率可分别高达77.22%、57.59%,实验结果表明,相比几种较为先进的人脸识别算法,该算法在处理单训练样本人脸识别问题时取得了更好的识别效果。For the issue that traditional face recognition methods could not get expected results in dealing with single sample per person （SSPP） because of lacking training images, this paper proposed the multiple manifold discriminative analysis （MM- DA） algorithm based on blocking clustering. Firstly, it divided each single training image into some non-overlapping local pat- ches with same size. Then,it clustered all local pitches on each manifold by cluster algorithm and used feature transform to maximize separabilities between classes. Finally, it computed the distances between testing manifold and manifolds of total training images and used nearest neighbor classifier to finish recognition. Experiments on ORL and FERET face databases verify the effectiveness and robustness of proposed algorithm. Experimental results show that proposed algorithm has better recogni- tion effect than several advanced algorithms in dealing with the SSPP problem.

关 键 词：人脸识别 单训练样本 多流形判别分析 子空间学习 分块聚类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]