基于结构化局部约束低秩表示的人脸识别  

Face recognition based on structured locality-constrained low rank representation

作　　者：蔡晓云 尹贺峰[1] CAI Xiaoyun;YIN Hefeng(School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China;Zhenjiang College, Zhenjiang, Jiangsu 212028, China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]镇江市高等专科学校,江苏镇江212028

出　　处：《江苏大学学报（自然科学版）》2020年第3期256-261,共6页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助项目(61672265);镇江市科技支撑计划项目(FZ2011034)。

摘　　要：针对传统的基于低秩表示的方法需要重新在字典矩阵上计算测试样本的表示系数,会导致计算复杂度升高,降低训练和测试样本表示系数之间的相关性等问题,提出了一种结构化局部约束低秩表示算法用于人脸识别的方法.在原始低秩表示中引入理想编码系数矩阵正则项,使训练样本的表示系数矩阵具有块对角结构;为保持数据的流形结构,引入局部约束项,使相似样本具有相似的表示系数;使用简单的线性分类器对测试样本进行分类.在AR,Extended Yale B,ORL和LFW这4个标准数据集上进行了试验结果验证.结果表明:该算法可以同时得到训练和测试样本的表示系数,对人脸图像中的遮挡、像素破坏和光照变化等具有鲁棒性.For the method of conventional low rank representation(LRR),the representation coefficients of test samples are calculated based on the learned dictionary,which results in high computational complexity and low correlation of representation coefficients between training and test samples.To solve the problems,a structured locality-constrained low rank representation(SLCLRR)method was proposed for face recognition.An ideal regularization term was introduced in LRR to encourage the representation matrix of training data with block-diagonal structure.A locality constraint was incorporated to acknowledge the intrinsic manifold structure of training data and make the similar samples with similar representations.Test samples were classified by a simple yet effective linear classifier.The verified experiments were conducted on the four benchmark datasets of AR,Extended Yale B,ORL and LFW.The results show that the proposed method can obtain the representation coefficients of training and test samples simultaneously with good robustness for occlusions,pixel corruptions and illumination variations in the face images.

关 键 词：人脸识别 低秩表示 块对角结构 局部约束项 线性分类器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]