低分辨率人脸图像的迭代标签传播识别算法  

Low-Resolution Face Recognition Based on Recursive Label Propagation Algorithm

作　　者：薛杉[1] 朱虹[1] 王婧[1] 史静[1] XUE Shah;ZHU Hong;WANG Jing;SHI Jing(Faculty of Automation and Information Engineering,Xi＇ an University of Technology,Xi＇an 710048)

机构地区：[1]西安理工大学自动化与信息工程学院,西安710048

出　　处：《模式识别与人工智能》2018年第7期602-611,共10页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61771386;61673318);湖北省教育厅科学技术研究项目(No.B2017080)资助~~

摘　　要：针对监控视频下低分辨率人脸识别中存在的特征表示能力不强及判别开集人脸图像身份不够准确的问题,文中提出低分辨率人脸图像的迭代标签传播识别算法.采用视觉几何组(VGG)提取人脸图像特征,依据特征的相似度获得高、低分辨率图像的映射关系.对已标记样本和未标记样本进行迭代标签传播,在每次迭代过程中,通过统计每个类别的置信度直方图,估计识别精确率逼近100%的自适应置信度阈值.根据阈值将确认的未标记样本更新至已标记样本集,提高算法召回率.在公共数据集上的实验表明,文中算法在精确率逼近100%的基础上,召回率取得较高值.In low-resolution face recognition, the feature representation ability is not robust and the discrimination result of open-set face recognition is inaccurate. Therefore, an algorithm for low-resolution face recognition based on reeursive label propagation algorithm is proposed. Firstly, VGG network is utilized to extract face representations. Secondly, the mapping relationship between high-resolution and low-resolution images can be acquired based on the similarity of feature vectors. Finally, the iteration label propagation algorithm is conducted on the labeled and unlabeled facial samples. During the iterations, the adaptive confidence threshold approaching to 100% recognition accuracy is estimated according to the confidence histogram of eaeh eategory. The identified unlabeled samples are updated to the labeled sample set based on the threshold, thereby the recognition recall rate is improved. Experimental results on public face datasets show that the proposed algorithm achieves a high recall rate with 100% precision.

关 键 词：标签传播 卷积神经网络(CNN) 低分辨率 人脸识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]