人脸检测中基于自适应ICA的特征提取算法  

Feature Extraction Based on Self-Adaptive ICA in Face Detection

作　　者：赵伟达[1] 张丽清[1] 

机构地区：[1]上海交通大学计算机科学与工程系,上海200240

出　　处：《计算机仿真》2007年第10期204-208,共5页Computer Simulation

摘　　要：如何从图片中提取出有效特征来区分人脸与非人脸一直是一个难题。文中提出了利用自适应独立成分分析（Self-Adaptive ICA）算法对图像结构信息非常敏感的特点,有效地从大量正面人脸图片中分离出人脸的局部特征,从而利用这些局部特征基底有效地表示人脸图片。自适应ICA算法的优点是能自适应的拟合图像数据的统计性质,而不用预先设定。通过比较待检测的人脸图片与非人脸图片在这组特征基底上的投影系数,可以较好的区分二者。实验结果也表明这种特征提取方法可以找到一组很好的人脸特征基底。使用这种方法构造的弱分类器的分类准确率在相同的误检率下比BoostedCascaded方法中的弱分类器高1%~1.5%。It is still a hard problem to extract efficient features from images to distinguish faces and non-face images.The paper presents a new approach to extract facial features from plenty of frontal face images with self-adaptive ICA algorithm which is sensitive to image structures so that face images can be represented efficiently by the projections on these facial features.It is an advantage of self-adaptive ICA algorithm to estimate image statistics without any assumptions in advance.Face and non-face images can be classified well via comparing their projections on the facial features.Computer simulations show that a set of good facial features can be found by the proposed approach.The comparison with Boosted Cascaded method also shows that weak classifiers employing the proposed approach yield 1% ～ 1.5% higher classification performances.

关 键 词：人脸检测 独立成分分析 特征提取 特征选择 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]