第16卷第2期 智能系统学报 Vol.16 No.2 2021年3月 CAAI Transactions on Intelligent Systems Mar.2021 D0:10.11992/tis.201910028 网络出版地址:https:/ns.cnki.net/kcms/detail/23.1538.TP.20200728.1555.008.html 稀疏综合字典学习的小样本人脸识别 狄岚2,矫慧文',梁久祯3 (1.江南大学人工智能与计算机学院,江苏无锡214122:2.道路交通安全公安部重点实验室,江苏无锡 214151:3.常州大学信息科学与工程学院,江苏常州213164) 摘要:传统以字典学习为基础的小样本人脸识别方法存在字典低辨别性、弱鲁棒性等缺点,对此,本文提出 稀疏综合字典学习模型。该模型有效利用和生成人脸变化,以镜像原理及Fisher准则扩充训练样本多样性,通 过构造混合特色字典、扩充干扰字典以及低秩字典原子,提取不同类别数据之间的共性、特殊性和异常情况, 从而提高算法识别率以及对表情变化、姿态变化、遮挡等异常情况的处理能力。在AR、YALEB、LFW等人脸 数据库进行仿真实验,实验结果验证了算法的有效性和可行性。 关键词:综合字典学习;人脸识别:类别特色字典;Fisher准则;小样本;图像扩充;镜像准则:扩充干扰字典;混 合特色字典;低秩字典 中图分类号:TP394文献标志码:A文章编号:1673-4785(2021)02-0218-10 中文引用格式:狄岚,矫慧文,梁久桢.稀疏综合字典学习的小样本人脸识别J.智能系统学报,2021,16(2):218-227. 英文引用格式:DI Lan,JIAO Huiwen,LIANG Jiuzhen..Sparse comprehensive dictionary learning for small--sample face recogni-- tionJI.CAAI transactions on intelligent systems,2021,16(2):218-227. Sparse comprehensive dictionary learning for small-sample face recognition DI Lan'2,JIAO Huiwen',LIANG Jiuzhen (1.School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi 214122,China;2.Laboratory of Ministry of Public Security for Road Traffic Safety,Wuxi 214151,China;3.School of Information Science and Engineering,Changzhou Uni- versity,Changzhou 213164,China) Abstract:Traditional small-sample face recognition methods based on dictionary learning have disadvantages such as poor dictionary discrimination and lack of robustness.In this paper,we propose a sparse comprehensive dictionary learning model.This model effectively utilizes and generates facial changes,expands the diversity of training samples by the mirror principle and Fisher's criterion,and extracts the commonalities,specialties,and anomalies between differ- ent categories of data by constructing a hybrid feature dictionary,extended interference dictionary,and low-rank diction- ary atoms.This strategy improves the recognition rate of the algorithm and its ability to handle abnormal situations such as expression changes,pose changes,and occlusions.The results of simulation experiments performed on the face data- bases AR,YALEB,and LFW verify the effectiveness and feasibility of the proposed algorithm. Keywords:comprehensive dictionary learning;face recognition;class-specific dictionary learning;Fisher discrimina- tion criterion;small sample;image expansion;mirror principle;extended interference dictionary;hybrid feature diction- ary:low-rank dictionary 近年来,深度学习进入蓬勃发展时代,以深度 收稿日期:2019-10-23.网络出版日期:2020-07-28. 基金项目:江苏省研究生科研与实践创新计划项目 学习为基础的图像识别)虽然识别准确率高,但 (KYCX19_1895,道路交通安全公安部重点实验室开 往往对硬件设备要求严格,具有训练时间长达数 放课题(2020 ZDSYSKFKT03-2,A类). 通信作者:狄岚.E-mail:dilan(@jiangnan.edu.cn. 周、样本量需求过大等不足之处。与之相比,基DOI: 10.11992/tis.201910028 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20200728.1555.008.html 稀疏综合字典学习的小样本人脸识别 狄岚1,2,矫慧文1 ,梁久祯3 (1. 江南大学 人工智能与计算机学院,江苏 无锡 214122; 2. 道路交通安全公安部重点实验室,江苏 无锡 214151; 3. 常州大学 信息科学与工程学院,江苏 常州 213164) 摘 要:传统以字典学习为基础的小样本人脸识别方法存在字典低辨别性、弱鲁棒性等缺点,对此,本文提出 稀疏综合字典学习模型。该模型有效利用和生成人脸变化,以镜像原理及 Fisher 准则扩充训练样本多样性,通 过构造混合特色字典、扩充干扰字典以及低秩字典原子,提取不同类别数据之间的共性、特殊性和异常情况, 从而提高算法识别率以及对表情变化、姿态变化、遮挡等异常情况的处理能力。在 AR、YALEB、LFW 等人脸 数据库进行仿真实验,实验结果验证了算法的有效性和可行性。 关键词:综合字典学习;人脸识别;类别特色字典;Fisher 准则;小样本;图像扩充;镜像准则;扩充干扰字典;混 合特色字典;低秩字典 中图分类号:TP394 文献标志码:A 文章编号:1673−4785(2021)02−0218−10 中文引用格式:狄岚, 矫慧文, 梁久祯. 稀疏综合字典学习的小样本人脸识别 [J]. 智能系统学报, 2021, 16(2): 218–227. 英文引用格式:DI Lan, JIAO Huiwen, LIANG Jiuzhen. Sparse comprehensive dictionary learning for small-sample face recognition[J]. CAAI transactions on intelligent systems, 2021, 16(2): 218–227. Sparse comprehensive dictionary learning for small-sample face recognition DI Lan1,2 ,JIAO Huiwen1 ,LIANG Jiuzhen3 (1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China; 2. Laboratory of Ministry of Public Security for Road Traffic Safety, Wuxi 214151, China; 3. School of Information Science and Engineering, Changzhou University, Changzhou 213164, China) Abstract: Traditional small-sample face recognition methods based on dictionary learning have disadvantages such as poor dictionary discrimination and lack of robustness. In this paper, we propose a sparse comprehensive dictionary learning model. This model effectively utilizes and generates facial changes, expands the diversity of training samples by the mirror principle and Fisher's criterion, and extracts the commonalities, specialties, and anomalies between different categories of data by constructing a hybrid feature dictionary, extended interference dictionary, and low-rank dictionary atoms. This strategy improves the recognition rate of the algorithm and its ability to handle abnormal situations such as expression changes, pose changes, and occlusions. The results of simulation experiments performed on the face databases AR, YALEB, and LFW verify the effectiveness and feasibility of the proposed algorithm. Keywords: comprehensive dictionary learning; face recognition; class-specific dictionary learning; Fisher discrimination criterion; small sample; image expansion; mirror principle; extended interference dictionary; hybrid feature dictionary; low-rank dictionary 近年来,深度学习进入蓬勃发展时代,以深度 学习为基础的图像识别[1-3] 虽然识别准确率高,但 往往对硬件设备要求严格,具有训练时间长达数 周、样本量需求过大等不足之处。与之相比,基 收稿日期:2019−10−23. 网络出版日期:2020−07−28. 基金项目:江苏省研究生科研与实践创新计划项 目 (KYCX19_1895); 道路交通安全公安部重点实验室开 放课题 (2020ZDSYSKFKT03-2, A 类). 通信作者:狄岚. E-mail: dilan@jiangnan.edu.cn. 第 16 卷第 2 期 智 能 系 统 学 报 Vol.16 No.2 2021 年 3 月 CAAI Transactions on Intelligent Systems Mar. 2021