【作者】 李云峰；

【导师】 欧宗瑛；

【作者基本信息】 大连理工大学 ， 机械设计及理论， 2006， 博士

【摘要】 人脸识别是一项极具有发展潜力的生物特征识别技术,研究人脸识别技术具有十分重要的理论和应用价值。最近几年,人脸识别技术取得了前所未有的发展,但其在实际应用中的识别精度仍然难以满足人们的预期要求,特别是采集图像中存在光照变化、摄像方位变异以及其它干扰时。识别系统采集的原始人脸图像通常以网格像素的灰度值集表示。孤立的像素灰度集合不能直接反映人脸的内蕴特征,引入适当的变换,将其映射到特征空间进行识别处理是行之有效的提高识别处理性能的途径。二维Gabor小波变换能够将相邻区域的像素联系起来,从不同的频率尺度和方向反映局部范围内图像像素灰度值的变化。二维Gabor小波变换系数描述了图像上各给定位置附近区域的灰度特征,在人脸图像的二维Gabor小波变换系数的基础上进行的特征提取和分类识别,称为基于Gabor小波变换的人脸识别。本文深入研究了利用二维Gabor小波变换进行人脸识别的理论方法和技术,论文的主要工作和贡献如下: (1) 本文对当前常用的人脸识别理论方法做了扼要的概括总结,结合最近几年国际上举办的一系列人脸识别评测活动,对当前人脸识别技术的研究现状、存在的问题和技术发展进行了论述。 (2) 研究了人脸图像的预处理。人脸图像的预处理就是将由图像采集设备采集到的人脸图像调整成规范化的图像,主要包括三个环节:人脸检测和眼睛定位、几何规范化、灰度规范化。论文重点研究了基于AdaBoost统计学习的人脸检测方法。 (3) 研究了二维Gabor小波变换及其在识别应用中的响应特性。二维Gabor小波变换是通过计算一组二维Gabor滤波器与图像上给定位置附近区域像素灰度值的卷积来实现的。二维Gabor小波是哺乳动物视觉皮层简单细胞接受场模型的良好近似。本文通过计算结果验证了可以通过选择Gabor滤波器的参数来表示人脸图像的局部特征,并且这种表示具有对摄像环境亮度的绝对水平变化不敏感的优点。基于二维Gabor小波变换进行的识别处理优于直接按原图像灰度的识别处理。 (4) 改进了经典的弹性束图匹配算法。弹性束图匹配算法采用标号图来表示人脸图像,标号图的节点用一组描述人脸局部特征的二维Gabor小波变换系数标示,这些节点位于人脸图像上对识别有意义的特征点位置上;标号图的边用描述相邻两个节点相对位置的度量信息来标示,由各边组成的网格图结构描述了整个人脸的几何特征。经典的弹性束图匹配算法首先将人脸图像与某一预定义人脸束图(即某一复合标号图,其节点为更多还原

【Abstract】 Face recognition is a biometric technology possessing great developable potential, researching on the face recognition technology has great theoretical and practical values. In recent years, face recognition technology has achieved unprecedented progress, but its recognition precision in practical applications still cannot satisfy the expectant demands of people, especially under the condition that variations of illumination, photographing azimuth or other disturbance exist in the image. The original face image captured by the recognition system usually is denoted by the grey values of grid pixels. Isolated grey values of pixels cannot reflect the characteristics contained in human face directly, mapping them into the feature space to recognize through adopting appropriate transform is an effective approach to improving the recognition performance. Two-dimensional Gabor wavelets transform can link the pixels in an adjacent region together, and reflect the changes of the grey values of pixels in a local area of an image from different frequency scales and orientations. Two-dimensional Gabor wavelets transform coefficients describe a small patch of grey values in an image around every given position, feature extraction and classification which are based on the two-dimensional Gabor wavelets transform coefficients of face image are called Gabor wavelets transform based face recognition. This dissertation researches into the theory and technology of face recognition through two-dimensional Gabor wavelets transform, the main work and contributions of the dissertation are as follows:(1) The commonly used face recognition theories and methods are summarized compendiously in this dissertation. The research actualities, existing problems and technology development of current face recognition technology are discussed based on a series of international face recognition evaluations which were conducted in recent years.(2) The preprocessing of face image is researched. The aim of face image preprocessing is to regularize the face image which is captured by image collecting devices to normalized image, it includes three steps mainly: face detection and eyes location, geometry normalization, grey value normalization. This dissertation emphasizes on the research of face detection method which is based on AdaBoost statistical learning.(3) Two-dimensional Gabor wavelets transform and its response characteristics in recognition applications are researched. Two-dimensional Gabor wavelets transform is realizedby computing the convolutions of a bank of two-dimensional Gabor filters and the grey values of pixels in an area around a given position in an image. Two-dimensional Gabor wavelets seem to be a good approximation to the receptive fields of the simple cells in the visual cortex of mammalians. In this dissertation, it is validated by the computational results that the local features of face images can be represented through selecting the parameters of Gabor filters, and this kind of representation has the merit of insensitiveness to the absolute brightness of the capturing environment. Recognition based on two-dimensional Gabor wavelets transform surpasses the one based on the grey values of the original image directly.(4) Classical elastic bunch graph matching algorithm is improved. Elastic bunch graph matching algorithm uses labeled graph to represent face image, every node of the labeled graph is labeled with a set of two-dimensional Gabor wavelets transform coefficients which describe the local facial feature, and these nodes lie at the feature point positions of the face image which is useful for recognition;every edge of the labeled graph is labeled with metric information on the relative position of two adjacent nodes, grid structure that is composed by all edges describes the geometrical feature of the whole face. The classical elastic bunch graph matching algorithm matches the face image to a predefined face bunch graph (namely a composite labeled graph, its every node is labeled with a set of Gabor wavelets transform coefficients of corresponding node of many labeled graph, its every edge is labeled with the average of corresponding edge of many labeled graph) in order to obtain the rough positions of the feature points firstly;then, every feature point is located through elastic fine adjustment;lastly, the two-dimensional Gabor wavelets transform coefficients are computed at the feature points and these coefficients are used for face classification and recognition. The computation of classical elastic bunch graph matching algorithm is prohibitive, in this dissertation, seven representative grid structures are obtained through the clustering of the grid structures of the face labeled graphs of many training images, these grid structures are used to constitute a template bunch of the face bunch graph. During the matching stage, the elastic bunch graph matching algorithm is combined with AdaBoost learning algorithm: firstly, the eyes are located by AdaBoost learning algorithm, using eye coordinates as the datum marks, the input image is geometrically normalized;then, the most appropriate grid structure is selected from the template bunch to determine the geometrical feature of a face, and precise matching is performed further based on the outcome. The matching computation is simplified largely after improvements.(5) The influence law of facial features which are represented by two-dimensional Gabor wavelets transform coefficients on face recognition and the corresponding solving methods are researched. This problem is researched from two angles: firstly, the two-dimensional Gaborwavelets transform coefficients at a given facial position are regarded as a local feature cell, the trace of the product of the inversion matrix of the within-class scatter matrix and the between-class scatter matrix is used as a criterion for weighing the discriminative capability of a local feature. By using this criterion, the contributions of the local features that lie at different facial positions to face recognition are analyzed. According to the contributions of different local features to face recognition, face recognition is implemented by the method of local feature weighting;then, a criterion for weighing the discriminative capability of single Gabor feature is proposed. By using this criterion, the discriminative capability of every Gabor wavelets feature of face image is analyzed, and the discriminative capability is linked to its position, frequency and orientation. These analyses can provide the evidences for the optimal selection of the positions of the facial feature points and the parameters of the Gabor wavelets filter bank. According to the discriminative capabilities of Gabor features, one can select the parts that are most advantageous to classification for face recognition.(6) Support Vector Machines classification algorithm is researched, a hierarchical decomposed Support Vector Machines binary decision tree classification scheme is proposed. The pattern set is divided into two parts with similar number of categories at every classification node of the decision tree. During the training stage, as for every classification node, the best partition which divides a pattern set into two parts is found by clustering firstly;then, these two parts are used for Support Vector Machines training. During the recognition stage, the pattern need to be classified is inputted from the root node of the decision tree, and its class label is determined by the leaf node. The classification times of Support Vector Machines in solving multiclass problem can be reduced largely by using this classification scheme.更多还原