变化字典学习与显著特征提取的单样本人脸识别  

Single Sample Face Recognition via Variant Dictionary Learning and Salient Feature Extraction

作　　者：王念兵[1] 吴秦[1] 梁久祯[2] 许洁[1] 张淮[1] 

机构地区：[1]江南大学物联网工程学院物联网技术应用教育部工程研究中心智能系统与网络计算研究所,江苏无锡214122 [2]常州大学信息科学与工程学院,江苏常州213164

出　　处：《小型微型计算机系统》2017年第9期2134-2138,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61202312;61673193)资助;中央高校基本科研业务费项目(JUSRP51510;JUSRP51635B)资助

摘　　要：针对单样本问题,基于不同的人脸区域在识别过程中的重要性不同这一事实,提出将能显著区分其它类的人脸区域作为提取的显著特征,并视为有较大区分度的块,剩下的区域视为普通块;再根据不同组中每块的稀疏表示重构残差给予不同的权重,以抑制普通块的影响同时保持有区分度的块的作用.为了减小人脸之间未对齐的影响,将每块训练图像对应的8邻域增加到训练集中,以实现样本的扩充;同时提出新的类内变化字典学习方法,学习得到共享的类内变化字典,以减小测试人脸未知变化的影响.文章的方法可以有效减小人脸局部信息缺失造成的影响,使得在AR、Extended Yale B、CMU-PIE人脸库上的表现超过其它单样本识别相关的方法,取得了最好的识别效果.Single sample per person makes face recognition much more difficult.According to the fact that different parts of faces have different importance on the process of face recognition,facial patches distinguished from other class are regarded as salient feature and grouped into more discriminative patches （MDP）,and others are assigned to less discriminative patches （LDP）.Blocks-weighted depended by the sparse representation residual of each patch will decrease the impact of LDP and maintain the effect of MDP.The 8 closet neighboring patches are extracted to address the misalignment of a patch.We propose to learn the intra-class variance dictionary so that the facial variance can be represented well.Our method can decrease the impact of regional information missing.Experimental results on the AR,Extended Yale B,CMU-PIE datasets show our method performs well than other specially designed single sample per person related face representation methods and achieves the best performance.

关 键 词：单样本 显著特征 稀疏表示 分块加权 类内变化字典 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]