摘要
实际环境(如监控)中常遇到大量低分辨率人脸图像需要识别.对低分辨率人脸的识别相对高分辨率更难,因其含有相对有限的判别信息.为此,通过在人脸识别(系统)构建阶段引入与低分辨率人脸相配对的高分辨率人脸,以提高识别性能成为最近研究的焦点之一.但这些研究仍存在以下不足:1)均要求高、低分辨率人脸样本间的全配对;2)识别系统构建时未利用给出的类信息,导致系统性能受限.事实上常常面对的应用场景是仅能获取部分配对和部分标号的高、低分辨率人脸样本集,即所谓的半配对半监督场景,对此提出一种用于低分辨率人脸识别的半配对半监督算法,以弥补现有相关研究的不足.在Yale和AR人脸数据集上的实验结果验证了该算法的有效性.
In the real environment, such as surveillance circumstances, there are a large number of low-resolution （LR） faces which are needed to be recognized. Compared with high-resolution （HR） face, LR has less discriminative details, so its recognition is more difficult. In order to improve the LR face recognition accuracy, the construction of LR face recognition system use not only the LR faces but also the HR faces corresponding to the LR faces in recent research. But there are two deficiencies in them： 1） HR faces and LR faces are required to be all paired; 2） the construction of face recognition system does not utilize any class information. Actually, it is the fact that HR faces and LR faces are always partially paired （semi-paried） and their class labels are partially known （semi supervised）. As a result, a semi-paired and semi-supervised algorithm for LR face recognition is developed to overcome the deficiencies of the relevant research. For the sake of utilizing the semipaired and semi supervised data more effectiviely, the implementation of the algorithm is divided into two stages. One stage is semi-paired learning and the other stage is semi-supervised learning. Promising experiments results on the Yale and AR face databases show the feasibility and effectiveness of the proposed method.
出处
《计算机研究与发展》
EI
CSCD
北大核心
2012年第11期2328-2333,共6页
Journal of Computer Research and Development
基金
国家自然科学基金项目(61170151)
江苏省高校自然科学研究项目(12KJB520018)
南京航空航天大学研究基金项目(NP2011030)
关键词
低分辨率人脸
高分辨率人脸
人脸识别
半配对学习
半监督学习
low-resolution face
high-resolution face
face recognition
semipaired learning
semisupervised learning
