电影视频人像检索系统  

A Face Retrieval System for Feature-Length Film

作　　者：陈立珍[1] 崔国勤[1] 李卓[1] 

机构地区：[1]中国科学院计算技术研究所前瞻研究中心

出　　处：《计算机仿真》2008年第6期187-190,236,共5页Computer Simulation

基　　金：国家自然科学基金项目(60473043);北京市科技新星计划项目(2005B54)

摘　　要：针对电影视频中序列特征,提出了一种人脸检索的方法及在电影视频中的应用。首先,运用训练好的Cascade AdaBoost人脸检测器选检测视频序列中的人脸图像,将检测到的人脸图像经标准化处理后投影到增量特征人脸空间中,得到人脸图像的向量表述,应用单类支持向量机进行训练和分类,同时根据分类的结果动态调整增量特征人脸空间方向和分类器的最佳分类超平面,实现对电影视频中特定演员的检索功能,对电影《小花》等的实验表明,这个方法可以得到理想的检索结果。Matching people based on their face images is a hard work because of the well - known problems of illumination, pose, size and expression variation. Fortunately, films have the happy benefit of series character that can easily be used to help retrieval. The objective of this work was automatic detection and identification of individuals in unconstrained consumer video, using Adaboost to select the number of frames per video shot in order to detect human faces first, then normalized the detected human faces and projected them to eigenspace. Using the face as the template, the paper generated one - class SVM, and got classifier for the faces that were detected. The experiment of the famous Chinese film＂ Little Flower＂ showed that this algorithm could get good retrieval results.

关 键 词：人脸检测 人脸识别 主成分分析 支持向量 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]