一种基于相关反馈的视频人脸算法

Novel face recognition relevance feedback algorithm for video

在保局投影算法(LPP)及支持向量机(SVM)的基础上提出了一种基于相关反馈的视频人脸识别算法.该算法通过合理的数据建模提取出视频中的时空连续性语义信息,同时能够发现人脸数据中内在的非线性结构信息而获得低维本质的流形结构,还能通过反馈学习来增加样本的标记类别.在UCSD/Honda视频人脸数据库和自采集数据库上进行比较的实验结果表明,该算法能够获得更好的识别效果.

How to fully utilize both spatial and temporal information in video to overcome the difficulties existing in the video-based face recognition, such as the low resolution of face images in video, large variations of face scale, radical changes of illumination and pose as well as occasional occlusion of different parts of faces, is the key problem. In this paper, on the basis of Locality Preserving Projections （LPP）, we propose a novel relevance feedback video face recognition method （RFVLPP）, which can preserve more spatial and temporal information hidden in the video face sequence using clustering, and make full use of the intrinsic nonlinear structure information to extract discriminative manifold features. The experiment compares RFVLPP with other algorithms on UCSD/Honda Video Database and our own Video Database. Experimental results show that the proposed approach can outperform state-of-the-art solutions for video- based face recognition.