摘要
提出了一种基于连续Adaboost算法的多视角人脸检测方法.人脸按其三维姿态被划分成若干个视点子类,针对每个子类使用Haar型特征设计了具有连续致信度输出的查找表型弱分类器形式,构造出弱分类器空间,采用连续Adaboost算法学习出基于视图的瀑布型人脸检测器.为了提高检测速度,使用了多分辨率搜索和姿态预估计策略.对于正面人脸检测,在CMU+MIT的正面人脸测试集合上检测的正确率为94.5%,误报57个;对于多视角人脸检测,在CMU侧面人脸测试集合上检测的正确率为89.8%,误报221个.在一台PentiumⅣ2.4GHz的PC上,处理一幅大小为320×240的图片平均需80ms.实验结果表明该方法十分有效,具有明显的应用价值.
In this paper, a multi-view face detection method based on real Adaboost algorithm is presented. Human faces are divided into several viewpoint categories according to their poses in 3D, and for each of these categories a form of weak classifiers in look-up-table （LUT） type is designed using Haar-like features that have confidences in real values as their outputs, and correspondingly its space of weak classifiers is constructed, from which the cascade face detector is learnt by using real Adaboost algorithm. For speed up, multi-resolution searching and pose prediction strategies are introduced. For frontal face detection, the experiments on CMU ＋ MIT frontal face test set result in a correct rate of 94.5 % with 57 false alarms; for multi-view face detection, the experiments on CMU profile face test set result in a correct rate of 89.8% with 221 false alarms. The average processing time on a PⅣ 2.4GHz PC is about 80 ms for a 320×240- pixel image. It can be seen that the proposed method is very efficient and has significant value in application.
出处
《计算机研究与发展》
EI
CSCD
北大核心
2005年第9期1612-1621,共10页
Journal of Computer Research and Development
基金
国家自然科学基金重点项目(60332010)
日本欧姆龙公司合作基金项目(0302J05)~~
关键词
Haar型特征
查找表型弱分类器
姿态估计
multi-view face detection
real Adaboost
Haar-like feature
LUT weak classifier
pose estimation
