摘要
针对人脸识别中存在的遮挡、阴影、反光等不同程度的数据破坏以及训练样本不充足导致识别率低的问题,提出一种基于隐式低秩表示联合稀疏表示(LatLRR_SRC,Latent Low-Rank Representation Sparse Representation Classification)的人脸识别方法.该方法首先采用隐式低秩表示(LatLRR,Latent Low-Rank Representation)算法将训练样本矩阵分解为两个低秩逼近矩阵和一个稀疏误差矩阵.然后将低秩逼近矩阵和稀疏误差矩阵联合构成完备字典,并用K-SVD算法对字典进行学习,得到测试样本在学习后字典下的稀疏表示.最后对测试样本利用上述隐式低秩表示分解的三部分的稀疏逼近计算残差,完成测试样本的分类识别.在Extend YaleB和CMU PIE人脸数据上的实验结果表明,基于LatLRR_SRC的人脸识别方法具有较高的识别率和稳定性.
In consideration of the occlusions,cast shadows,specularities,corruptions of different level in the images and insufficient samples,resulting the lower recognition rate,a face recognition on method of joint spare representation based on Latent Low-Rank Representation(LatLRR_SRC)is proposed.Firstly,in the method,training samples are decomposed into two low rank approximation matrix and a sparse error matrix using Latent Low-Rank Representation(LatLRR)algorithm.Then,Constructing the dictionary by using both the low rank approximation matrix and the error matrix,on the basis of this,training dictionary by using K-SVD algorithm.Finally,the test samples can be sparse-represented in the dictionary by using the sparse approximation of this three parts calculates the residual which used for classification.On the face database of Extend YaleB and CMU PIE,experimental results show that face recognition method proposed in this paper has a higher recognition rate and stability.
出处
《云南师范大学学报(自然科学版)》
2017年第1期43-51,共9页
Journal of Yunnan Normal University:Natural Sciences Edition
基金
云南省教育厅科学研究基金资助项目(2014Y145)
云南省哲学社会科学规划项目资助(QN2015067)
云南师范大学博士科研启动基金资助项目(01000205020503064)
关键词
隐式低秩表示
稀疏表示
完备字典
人脸识别
Latent Low-Rank representation
Sparse representation
Completed dictionary
Face recognition