基于核扩展混合块字典的单样本人脸识别研究  

Face Recognition with a Single Training Sample Per Person Based on Kernel Extended Hybrid Block Dictionary Learning

作　　者：马杲东 吕非 童莹[3] 曹雪虹[3] MA Gao-dong;LYU Fei;TONG Ying;CAO Xue-hong(School of Telecommunications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;NARI Group Corporation,Nanjing 211106,China;School of Information and Communication Engineering,Nanjing Institute of Technology,Nanjing 211167,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南瑞集团有限公司,江苏南京211106 [3]南京工程学院信息与通信工程学院,江苏南京211167

出　　处：《计算机技术与发展》2022年第1期104-110,116,共8页Computer Technology and Development

基　　金：国家自然科学基金(61703201);江苏省自然科学基金(BK20170765)。

摘　　要：稀疏表示分类(sparse representation-based classification,SRC)在样本数量充足下的人脸识别中具有较好的识别效果。然而由于基本字典缺乏判别性同时过度依赖于字典中每类样本的原子数目,稀疏表示分类在真实情况下的单样本(每类样本只有一张训练样本)人脸识别任务中缺乏鲁棒性。针对以上问题,该文提出了基于核扩展混合块字典的单样本人脸识别方法。首先,对样本进行分块处理,分别对分块图像进行核判别分析(kernel discriminant analysis,KDA)投影降维,提取图像的局部特征信息构成更具判别性的基本块字典;然后,为经过KDA投影之后的分块样本分别构建遮挡字典和类内差异字典来描述样本中的大面积连续遮挡以及光照、表情等类内差异信息,将遮挡字典和类内差异字典共同组合成混合块字典,使混合块字典能够更好地描述测试样本中不同类型的差异信息;最后,将测试样本表示为基本块字典和混合块字典的稀疏线性组合,根据重构残差进行分类识别,从而实现真实情况下的单样本人脸识别。在标准人脸库CAS-PEAL,AR以及真实人脸库LFW和PubFig上的实验结果表明,该方法与其他方法相比有较好的结果。Sparse representation-based classification(SRC)is effective in face recognition with sufficient samples.However,due to the lack of discriminativeness of its basic dictionary and excessive dependence on the number of atoms of each class of sample in the dictionary,SRC lacks robustness in face recognition tasks of single sample per person(SSPP).Therefore,we propose a single sample face recognition method based on kernel extended hybrid block dictionary.Firstly,the samples are divided into blocks,and the kernel discriminant analysis(KDA)projection dimension reduction is performed on the divided images respectively,and the local feature information of the image is extracted to form a more discriminative basic block dictionary.Then,for the blocked samples after KDA projection,an occlusion dictionary and an intra-class difference dictionary are constructed to describe the large-area continuous occlusion in the sample,as well as the intra-class difference information such as illumination and expression.The occlusion dictionary and the intra-class dictionary are combined to construct the hybrid block dictionary and enable the hybrid block dictionary to better describe the intra-class variation in the test sample.Finally,the test samples are represented as sparse linear combinations of the basic block dictionary and the hybrid block dictionary.The classification is determined by the reconstruction residuals.Experiment on standard face databases CAS-PEAL,AR and real face databases LFW and PubFig shows that the proposed method has better results compared with other methods.

关 键 词：稀疏表示分类 核判别分析 人脸识别 混合块字典 单样本 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]