摘要
遮挡人脸识别是人脸识别系统面临的挑战之一。在自然场景下,人脸特征通常被口罩等物品遮挡,导致人脸特征不完整,从而无法正确提取人脸特征信息,严重影响最终的识别结果。针对有遮挡条件下人脸识别效果较差的问题,通过利用低秩技术和二进制标签松弛模型的优势,该文提出了一种新的基于二进制松弛标签的回归模型。该模型通过学习一个更松弛的标签矩阵来代替严格的0-1标签矩阵,从而扩大了样本之间的类间距离,同时对二进制松弛标签矩阵采用低秩约束,以提高样本的类内相似性。因此,该方法能够提取出更多具有判别性的特征,从而有利于遮挡条件下的人脸识别。此外,通过引入的正则化项,有效避免了该方法的过拟合问题。在Yale B、LFW和CMU PIE数据集上的实验结果表明,该方法不仅能在实验室环境下获得较高的识别率,在自然场景下仍然能取得较好的识别性能。
Occlusion face recognition is one of the challenges faced by face recognition systems.In natural scenes,face features are usually occluded by masks and other items,resulting in incomplete face features,which cannot correctly extract facial feature information and seriously affect the final recognition results.Aiming at the problem of poor face recognition under occlusion,we propose a new regression model based on binary label relaxation by using the advantages of low-rank technology and binary label relaxation model.The model expands the inter-class distance between samples by learning a more relaxed label matrix instead of the strict 0-1 label matrix,and uses low-rank constraints on the binary relaxed label matrix to improve the intra-class similarity of the samples.Therefore,the proposed method can extract more discriminative features,which is beneficial to face recognition under occlusion.In addition,the introduced regularization term effectively avoids the over-fitting problem of the proposed method.The experimental results on the Yale B,LFW and CMU PIE datasets show that the proposed method can not only obtain a higher recognition rate in the laboratory environment,but also achieve better recognition performance in the natural scene.
作者
韩肖
马祥
HAN Xiao;MA Xiang(School of Information Engineering,Chang’an University,Xi’an 710064,China)
出处
《计算机技术与发展》
2022年第1期1-6,共6页
Computer Technology and Development
基金
国家自然科学基金(61771075)
中央高校基本科研业务费资助项目(300102249203)。
关键词
人脸识别
低秩技术
二进制松弛标签
特征提取
遮挡
face recognition
low rank technology
binary relaxation label
feature extraction
occlusion
