基于Log-Gabor小波和二维半监督判别分析的人脸图像检索  

Log-Gabor and 2D semi-supervised discriminant analysis based face image retrieval

作　　者：黄荣兵[1] 郎方年[1] 施展[2] 

机构地区：[1]成都大学四川省高校模式识别与智能信息处理重点实验室,成都610106 [2]华南理工大学电子与信息学院,广州510640

出　　处：《计算机应用研究》2012年第1期393-396,共4页Application Research of Computers

基　　金：四川省杰出青年基金资助项目(2010JQ0032);成都大学校青年基金资助项目(2010XJZ07)

摘　　要：针对光照、表情变化给人脸识别造成的影响以及大型人脸图像库的训练样本中只有部分标记的问题,结合多通道Log-Gabor小波和半监督流形学习算法,提出一种新的人脸图像检索方法。该方法首先使用Log-Ga-bor小波对人脸图像进行滤波获得特征矩阵,进一步利用提出的二维半监督流形学习算法进行维数约简,得到低维判别特征。由于该方法直接作用于Log-Gabor特征矩阵,克服了小样本带来的奇异问题;另外,通过充分利用标记和未标记信息,还保留了数据的局部流形结构,增强了特征匹配的相似性。在CMU PIE和AR人脸数据库上的实验结果表明,该方法有效且优于其他方法。Face recognition methods are easily affected by variances of lighting and expression,and there are often only part of sample faces labeled in large image database.To address the two problems,this paper presented a face image retrieval method using multi-channel Log-Gabor wavelet based semi-supervised manifold learning algorithm.Firstly,used Log-Gabor to filter the face image to obtain feature matrix.Then proposed two-dimension semi-supervised manifold learning algorithm to extract the discriminative submanifolds.As the proposed method operated on Log-Gabor feature matrix directly,small sample problem was overcome.Furthermore,by making use of labeled and unlabeled information,preserved the local manifold structure of image data.So enhanced the similarity of feature matching.Experimental results on CMU PIE and AR databases show that the proposed approach is effective and superior to others.

关 键 词：LOG-GABOR小波 人脸识别 图像检索 半监督学习 流形学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]