基于RPCA与LDA算法融合的人脸遮挡识别方法研究  

Research on Face Occlusion Recognition Based on Fusion of RPCA and LDA

作　　者：杨国庆 张慧 杜敏 YANG Guoqing;ZHANG Hui;DU Min(School of Control and Mechanical Engineering,TCU,Tianjin 300384,China)

机构地区：[1]天津城建大学控制与机械工程学院,天津300384

出　　处：《天津城建大学学报》2022年第3期210-215,共6页Journal of Tianjin Chengjian University

基　　金：天津市企业科技特派员项目(19JCTPJC50600);天津市研究生科研创新项目(2020YJSZXS19)。

摘　　要：针对人脸遮挡后部分局部信息属性改变的问题,提出了一种基于鲁棒性主成分分析的全局特征向量和误差特征向量融合算法,该算法具较强的鲁棒性,在处理大面积遮挡的人脸图像时能恢复出更为准确的低秩矩阵.首先对采集的人脸图像从灰度化、补光、二值化、均衡化等方面进行图像预处理;其次将提取的人脸特征信息进行融合;最后对融合的特征人脸使用LDA算法进行特征向量提取和识别分析,并采用经典AR人脸数据库和ORL人脸库对该算法进行仿真实验.结果表明:融合的双属性算法在遮挡面积达40%的情况下仍具有79.23%的识别率,人脸遮挡识别率得到有效提高;同时进行了人脸遮挡误识率实验,实验表明最低误识率为0.05%,随着识别数量的增加,误识率趋于稳定、达到0.1%,且具有更优的实用性和适应性.To solve the problem of partial information attribute change after face occlusion,this paper proposes a global eigenvector and error eigenvector fusion algorithm based on robust principal component analysis.First,the face image is pre-processed from the aspects of Grayscale,light compensation,binarization and equalization.Secondly,the extracted face feature information is fused.Finally,the fused feature face is extracted and analyzed by LDA Algorithm,and the algorithm is simulated with classic AR face database and ORL face database.The results show that:The fused bi-attribute algorithm still has a recognition rate of 79.23%when the occlusion area reaches 40%,and the face occlusion recognition rate is effectively improved.At the same time,the experiment of face occlusion error recognition rate shows that the minimum error recognition rate is 0.05%.With the increase of recognition number,the false recognition rate tends to reach 0.1%,and the algorithm has better practicability and adaptability.

关 键 词：人脸遮挡识别 双属性 鲁棒性主成分分析 LDA算法 特征向量 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]