结构化加权稀疏低秩恢复算法在人脸识别中的应用  

A low rank recovery algorithm for face recognition with structured and weighted sparse constraint

作　　者：吴小艺 吴小俊[1] WU Xiaoyi;WU Xiaojun(School of Internet of Things Engineering,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学物联网工程学院

出　　处：《智能系统学报》2019年第3期455-463,共9页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61672265,61373055);江苏省教育厅科技成果产业化推进项目(JH10-28);江苏省产学研创新项目(BY2012059)

摘　　要：针对训练样本或测试样本存在污损的情况,提出一种结构化加权稀疏低秩恢复算法(structured andweighted-sparse low rank representation,SWLRR)。SWLRR对低秩表示进行加权稀疏约束和结构化约束,使得低秩表示系数更加趋近于块对角结构,进而可获得具有判别性的低秩表示。SWLRR将训练样本恢复成干净训练样本后,再根据原始训练样本和恢复后的训练样本学习到低秩投影矩阵,把测试样本投影到相应的低秩子空间,即可有效地去除测试样本中的污损部分。在几个人脸数据库上的实验结果验证了 SWLRR在不同情况下的有效性和鲁棒性。Herein,a structured and weighted sparse low-rank recovery algorithm (SWLRR) is proposed to deal with trained or tested samples that are corrupt.The SWLRR constrains the low-rank representation by incorporating the struc-tured and weighted sparse constraints,enabling the low-rank representation coefficient matrix to be closer to the block diagonal.Then,a discriminative structured representation can be obtained.After recovering the clean training samples from the corrupted training samples using SWLRR,the low-rank projection matrix is learnt by the low-rank projection matrix according to the original and recovered training samples,whereas the test samples are projected into the corres-ponding low-rank subspaces.In this way,the corrupted regions can be removed efficiently from the test samples.The experimental results on several face databases validate the effectiveness and robustness of the SWLRR under different situations.

关 键 词：人脸识别 结构化 加权稀疏 低秩表示 子空间投影 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]