摘要
针对稀疏表示分类(sparse representation-based classification,SRC)算法在噪声、遮挡或者光照变化等情况下面部图像识别率较差的问题,对SRC模型进行算法优化,将L1损失函数替代L2损失函数用以求解稀疏解,并且采用L1范数和L2范数对L1损失函数最小化问题进行正则化。在3个具有挑战性的人脸数据集中挑选不同的光照、表情和遮挡条件时的人脸图像,并适当地加入噪声,分析在不同数据条件下SRC优化模型的性能,进而研究正则化参数在数据样本与稀疏性之间的修正关系。实验结果表明:所提出的两种SRC优化模型在不同的数据库和样本间具有不一样的性能,L1损失函数与L1正则化的组合在噪声条件时表现突出,L1损失函数与L2正则化的组合在遮挡条件下具有更高的鲁棒性。
Aiming at the problem that the facial image recognition rate is poor in sparse representation-based classification( SRC) algorithms in noise,occlusion,or lighting changes,this article optimizes the SRC model to replace the L1 loss function with the L1 loss function. In order to solve the sparse solution, L1 norm and L2 norm are used to regularize the L1 loss function minimization problem. In this paper,face images under different lighting,expression and occlusion conditions are selected from three challenging face data sets,and the noise is appropriately added to analyze the performance of SRC optimization model under different data conditions,so as to study the correction relationship between regularization parameters and sparsity of data samples. The experimental results show that the two SRC optimization models proposed in this paper have different performances in different databases and samples. The combination of L1 loss function and L1 regularization is prominent in noise condition,and the combination of L1 loss function and L2 regularization is more robust in occlusion condition.
作者
吉朝明
宋铁成
JI Chaoming;SONG Tiecheng(Department of Information Engineering,Sichuan Vocational and Technical College of Communications,Chengdu 611130,China;Department of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)
出处
《重庆理工大学学报（自然科学）》
CAS
北大核心
2020年第2期120-126,共7页
Journal of Chongqing University of Technology：Natural Science
基金
国家自然科学基金资助项目(61702065)
四川省教育信息化应用与发展研究中心项目(JYXX18-030)
关键词
人脸识别
稀疏表示
优化算法
范数正则化
face recognition
sparse representation
optimization algorithm
norm regularization
