一种基于低秩恢复稀疏表示分类器的人脸识别方法  

Face Recognition Method Based on Low-rank Recovery Sparse Representation Classifier

作　　者：杜海顺[1] 张旭东[1] 侯彦东[1] 金勇[1] 

机构地区：[1]河南大学图像处理与模式识别研究所,开封475004

出　　处：《计算机科学》2014年第4期309-313,共5页Computer Science

基　　金：国家自然科学基金(U1204611);河南省科技厅基础与前沿技术研究计划项目(132300410474);河南省教育厅科学技术重点研究项目(12A520008)资助

摘　　要：针对基于稀疏表示分类器(Sparse Representation-based Classification,SRC)的人脸识别方法用单位阵作误差字典不能很好地描述人脸图像噪声和误差以及由于训练样本不足可能造成字典不完备的问题,提出一种基于低秩恢复稀疏表示分类器(Low Rank Recovery Sparse Representation-based Classification,LRR_SRC)的人脸识别方法。该方法首先采用低秩矩阵恢复(LRR)算法将训练样本矩阵分解为一个低秩逼近矩阵和一个稀疏误差矩阵。然后,由低秩逼近矩阵和误差矩阵组成字典。在此基础上,得到测试样本在该字典下的稀疏表示。更进一步,基于测试样本的稀疏表示系数和字典,对测试样本进行类关联重构,并计算其类关联重构误差。最后,基于类关联重构误差,完成测试样本的分类识别。在YaleB和CMU PIE人脸数据库上的实验结果表明,提出的基于LRR_SRC的人脸识别方法具有较高的识别率。A face recognition method based on low-rank recovery sparse representation classifier （LRR_SRC） was proposed to overcome the disadvantages of the face recognition of sparse representation-based classification （SRC）,including the poor performance of the unit matrix as the error dictionary in the progress of describing the noise and error of the face images,and the dictionary incompletion caused by the insufficiency of the training samples.Firstly,in this method,training samples are decomposed into a low rank approximation matrix and a sparse error matrix using low-rank recovery （LRR） algorithm.And then,the low-rank approximation matrix and the error matrix compose a dictionary.On the basis of this,the sparse representation of the given test sample can be obtained under this dictionary.Further,using the sparse coefficients associated with the special class,LRR_SRC can approximate the given test sample and calculate the reconstruction errer between the given test sample with its approximation associated with the special class.Based on the reconstruction error associated with special class,the given test sample can be classified accurately.Experimental results on face database of YaleB and CMU PIE show that face recognition method proposed in this paper has a higher recognition rats

关 键 词：低秩矩阵恢复 稀疏表示 误差字典 人脸识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]