【作者】 郭志强；

【导师】 杨杰；

【作者基本信息】 武汉理工大学 ， 通信与信息系统， 2010， 博士

【摘要】 人脸识别是当前生物特征识别中的研究热点,提取稳定可靠、区别于其它个体的特征是人脸识别的关键。其中基于子空间投影的特征提取方法,因其算法简单、识别高效而备受人们的亲睐。本文以人脸识别为目标,针对现有人脸图像特征提取存在的几个问题,以子空间分析方法为中心,多种特征提取手段相结合,采用特征提取的混合模型,得到更加可靠有效的识别特征,取得了如下成果。对奇异值分析的人脸识别方法进行了改进,给出了一种包含平均脸奇异值分解的线性鉴别人脸特征提取方法。首先选用练训样本的均值图像作为标准图像,把训练样本投影到标准图像经奇异值分解产生的基空间中,其次提取投影系数矩阵左上角信息作为初步特征,最后再采用LDA降维,提取最终的特征。该方法改善了奇异值分解用于人脸识别基空间不一致的缺陷,同时又增加了特征的类别信息,也避免了LDA的小样本问题。将该方法扩展到非线性方式,给出了包含平均脸奇异值分解的核线性鉴别分析特征提取方法,进一步提取人脸的非线性特征。实验证明,文中提出的方法在识别率上,优于现在有奇异值分解及其改进人脸识别方法。对二维特征人脸识别方法进行了研究,提出了横向2DPCA纵向2DLDA的、双向压缩投影的子空间人脸识别方法,该方法在进行一次2DPCA运算后,对特征矩阵进行转置,再进行2DLDA运算。与(2D)2PCA与(2D)2LDA相比,该方法充分利用了2DPCA和2DLDA的优点,既包含了样本的类别信息,又消除了图像矩阵行和列的相关性,有效地提取了行和列的识别信息,识别特征维数也大幅度减少,在ORL和FERET人脸库上实验表明了其有效性。文中对了分块多投影和分块双向多投影二维特征提取方法进行了研究。分块多投影特征提取方法,针对现有分块单投影特征提取方法中每一子图均采用相同投影矩阵,而对人脸局部信息不加以区别这一问题,构造了分块多投影矩阵,使不同的子图对应不同：的投影矩阵,有效地利甩了人脸局部信息,使识别率得到了提高,在ORL人脸库上实验表明；了其有效性。提出一种基于小波包和PCA变换相结合的特征级融合人脸识别方法,首先对人脸图像进行小波包分解,对分解后的低频子图进行PCA分解,得低频主分量,然后选取含有丰富人脸特征的高频子图进行加权融合,对融合后的高频子图再进行PCA分解,得高频主分量,最后对高低频主分量进行融合处理,得到最终的鉴别特征。分别在ORL和YaleA人脸库上进行试验,实验结果表明该方法提高了识别率。针对保局投影人脸识别方法进行了研究,提出了核判别保局投影算法,即KDLPP。该算法通过核技巧将人脸样本映射到高维空间,在高维空间中有效结合人脸局部的流形结构和人脸的判别信息构建了新的目标函数,其优点是在保持人脸流形结构的基础上,充分利用了样本的类别信息,并采用核方法提取了人脸的非线性特征。在ORL和UMIST人脸库上的实验表明,该方法的识别率整体优于LPP、DLPP和KLPP。针对二维保局投影只在图像的横方向进行数据压缩,提取的特征维数高的问题,首先给出了可选的二维保局投影(A2DLPP)方法,然后将二维保局投影和可选的二维保局投影结合,设计了双向压缩二维保局投影方法,即(2D)2LPP算法。该算法分别从横向和纵向两个方向施实2DLPP,使图像的横向和纵向的维数都得到有效的约简。实验结果表明,(2D)2LPP无论是从识别率还是识别时间上都优于2DLPP和A2DLPP。更多还原

【Abstract】 Face recognition is the current well acknowledged and hot issues in the field of biometric identification, the key of which is to extract stable, reliable, distinct features. Subspace analysis of the feature extraction is main method due to simplicity and efficiency. In this paper, face recognition as the target, in order to resolve several feature extraction issues, a variety of feature extraction methods based on, subspace analysis are proposed, which combines with feature extraction mixture model and gets more reliable and effective features. The contributions of the dissertation can be noted as following.A modified method to extract features of face image based on Singular Value Decomposition (SVD) and Linear Discriminant Analysis (LDA) is proposed. Firstly, the mean image of all train samples is selected as a standard face image, and all the train samples are projected into the two orthogonal matrixes which come from the SVD of the standard face image. Then the left-top informations of projecting coefficient matrix are extracted as the primary feature. Finally, LDA is manipulated to extract the recognition feature. In this method, the problem of the equivalent basis space with SVD used into face recognition is amended,at the same time,class information of samples is added and problem of small sample with LDA is abolished. A new method to extract featuress of face image based on Singular Value Decomposition (SVD) and Kernel Discriminant Analysis (KDA) is proposed. Non-linear feature is accordingly extracted.Experiments are conducted on-ORL and CAS-PEAL databases, the results of which have indicated the method is effectiveSubspace method combining.2DPCA with 2DLDA is proposed for face recognition.This method performs 2DLDA or 2DPCA twice, one is 2DPCA in horizontal direction and the other is 2DLDA in vertical direction. The,advantage of this-method over the standard (2D)2PCA and (2D)2LDA method is that the former not only includes the class information of samples but also eliminates the image matrix correlation of the row and column. At the same time,it seeks both the column and row information, the dimension also is less than standard 2DLDA and 2DPCA. Experiments on ORL and FERET database indicate that this method is effective. This paper targets the solution of all the sub-image choose the same mapping matrix and do not discriminate facial local information in Module single-projecting feature extraction method, Module multi-projecting feature extraction method constructs module multi-mapping matrix, making different sub-image projcet different matrix. In this way, the rat of recognition is enhanced by making use of local facial information effectively and we have done experiment in ORL face database and proved its effectivenessA method of feature fusion face recognition based on wavelet pack transform and principal component analysis is proposed. Firstly, each face image was decomposed into sixteen sub-image by using two-dimensional discrete wavelet transform, and then PCA was using to extract the feature of low frequency sub-image. The selection of the high frequency sub-images includes abundant of human face information to be combined. Then, PCA was using to extract the feature of high frequency fusion image. All the extracted features were further fused and used for face classification. The experiments on the ORL face database and YaleA face database indicate that the method can reach a higher recognition rate.A new face image feature extraction and recognition algorithm based on kernel discriminant locality presercing projections was proposed, in which face images are projected into high dimensional feature spaces by using kernel trick. Then, in kernel space a new objective function is constructed with the face manifold local structure information is combined with the labels’information. The advantage of the method is not only the face manifold is preserved, but also the label information is being used, at the same time, non-linear feature is abstracted. Experiments have been done on ORL and UMIST, the experimental results showed that KDLPP is powerful than LPP, DLPP and KLPP.Two-dimensional locality preserving projection (2DLPP) extracts features only in the horizontal direction of the image and feature’s dimension is high. An alternative two-dimensional locality preserving projection (A2DLPP) method is given, and then double two-dimensional locality preserving projection((2D)2LPP) is designed. The algorithm executes 2DLPP separately from both horizontal and vertical directions, so that the features of the horizontal and vertical dimensions have been an effective reduction. Experimental results show that, both the recognition rate and recognition time of (2D) 2LPP are superior to 2DLPP and A2DLPP.更多还原