【作者】 张立刚；

【导师】 何东健；

【作者基本信息】 西北农林科技大学 ， 计算机应用技术， 2008， 硕士

【摘要】 人脸检测与识别技术是生物特征鉴别技术中研究最多和最热门的技术之一,它已经在身份认证、安全检查、罪犯查询、人机交互等广泛领域得到了初步应用。在人脸检测研究中,构建快速而精确的检测方法一直是该领域的研究热点;在人脸识别研究中,如何克服获取图像光线、表情、视角等变化的影响,提高识别率则是迫切需要研究的问题。针对这两个问题,本文以彩色和灰色正面人脸静态图像为研究对象,将模式识别理论和图像处理技术相结合,重点研究基于LVQ人工神经网络(ANN)的肤色像素检测和基于模板匹配的人脸精确检测方法,以及基于小波包分解(WPD)和(2D)2PCA的不同变化条件人脸图像的识别方法,为建立快速精确的人脸识别系统提供技术依据。本文的主要研究工作如下:(1)针对现有人脸检测系统检测精度和速度不平衡的问题,提出了一种基于LVQ ANN的肤色检测与基于模板匹配的精确人脸检测相结合的方法。该方法在获取肤色像素基础上,采用基于全局搜索的Mosaic方法预定位人脸区域。以CVL人脸库图像实验结果表明,LVQ ANN实现了较满意的肤色像素检测效果,又能提高检测速度;Mosaic方法成功地实现了人脸区域的预定位。(2)为在预定位人脸区域中实现精确的人脸检测,采用一种基于模板匹配的人脸检测方法。该方法首先构建基于R分量的标准灰度人脸模板,然后以相关性系数为匹配准则,使用多尺寸人脸模板实现不同尺寸人脸的检测。实验结果表明,CVL人脸库中常态组、微笑组和大笑组的正确检测率分别为100%、100%和93.6%;与仅采用模板匹配法相比,检测速度从1870.6s/幅提高到208.4s/幅。(3)为解决从图像小波包分解得到节点图像中选取显著节点困难的问题,提出了采用(2D)2 PCA和最邻近分类器测试所有节点图像的正确识别率(CRR),并依据识别率选取出“成功”节点图像的方法。(4)为了有效组合“成功”节点的特征矩阵,提出了一种测量测试图像和库图像距离的方法。该方法以“成功”节点图像特征矩阵的加权距离和,做为测试图像和库图像的距离,既考虑了全局和局部特征,又考虑了不同节点图像的识别贡献率,人脸识别实验结果表明该测量方法有效地提高了识别率。(5)针对变化人脸图像识别困难的问题,提出了一种基于WPD和(2D)2PCA的人脸识别方法。首先,对图像进行小波包分解,采用(2D)2PCA和最邻近分类器得到子节点的正确识别率,选取具有较大识别率的节点作为“成功”节点,然后,组合“成功”节点的特征矩阵,计算测试图像与库图像的距离,最后,采用最邻近分类器实现识别。(6)以MATLAB 7.0为工具编程实现基于WPD和(2D)2PCA的人脸识别方法,并以CMU PIE、Yale和UMIST人脸库图像为测试对象,分别进行光照、表情和视角变化图像的识别性能实验,以原图像采用(2D)2PCA和最邻近分类器的识别率为对比标准,结果表明,本文方法在3个实验中的识别率均高于标准识别率,其中,光照变化时识别能力最好,最高识别率为98.795%;表情变化其次,最高为89.796%,视角变化最差,最高为36.047%。(7)实验表明,距离尺度和小波函数的选取对多分辨率下节点的识别率有较大影响。L1在主体节点上的识别率高,而L2在细节节点上的识别率高;小波函数对不同条件图像识别效果也各不相同。因此,要根据图像变化条件选取节点、距离尺度和小波函数。由试验提出了如下选取规则:光照变化时,采用L1和Daubechies4下的A1、A2、H2、V2、HH2组合;表情变化时,采用L1和Haar下的A2。(8)本文提出的方法在视角变化时效果并不理想,尚需研究并寻求其它特征提取方法。更多还原

【Abstract】 The technology of face detection and recognition is one of the most widely investigated technologies in the filed of Biometric Identification, and it has been used in such areas as identity authentication, security check-up, criminal enquiry, human-computer interaction etc.In regard to face detection, proposing a detection method with high speed and accuracy remains a research hot spot. As to face recognition, due to the great variations of illumination, expression, viewpoint, age, etc. of face images, obtaining high recognition rates under these conditions still is a difficult task and research focus point. With respect to these two problems, this dissertation takes colour and gray static frontal facial images as research objects, and studies face detection and recognition methods based on combination of pattern recognition theory and image processing technology. The main content includes skin pixel extraction method on the basis of LVQ artificial neural network, face detection method using template matching technology, and a novel face recognition method employing (2D)2PCA and WPD under varying illumination, expression and pose conditions. This research targets for providing technology supports for a high-speed and accurate face recognition system. The main contributions of this research include:(1) In order to solve the problem that detection speed and accuracy of current face detection system is unbalanced, a method that extracts skin pixels using LVQ ANN and detects face based on template matching is proposed. Firstly, An LVQ ANN is used to extract skin pixels. Then, a Mosaic method is prompted to primarily locate the face region through searching within the whole image. Experiments on images from CVL indicate that the LVQ ANN gains satisfactory extraction accuracy as well as high speed, and the Mosaic method could successfully pre-locate the face region.(2) A method using template matching is adopted to detect face in the pre-located face region. First of all, a gray standard face template is gained by using only R channel of RGB images.Then, face is detected in the pre-located face region using template matching by taking relativity coefficient as the matching rule. In the end,the location and size of this face are obtained. Experimental results of three testing sets ( normal, smile and big smile sets) from CVL database show that the adopted method obtains good detection accuracy as well as speed. In the concrete, it gains 100%,100% and 93.6% correct detection rates respectively. Meanwhile, its detection speed increases from 1870.6 second/image to 208.4 second/image comparied with only adopting template matching on the original image.(3) To address the difficult problem of choosing remarkable plots from all plots gained via WPD on the original image, a method that selects“successful”plots according to the correct recognition rates (CRRs) of plots is proposed. These CRRs are obtained by combining (2D)2PCA with the nearest neighborhood classifier.(4) Aiming at efficiently fusing the feature matrixes of“successful”plots, a distance measurement between testing image and database image is presented. The L1 or L2 distances between feature matrixes of selected“successful”plots of testing image and each database image are calculated, and then taking the weighted sum of these distances as final distance. This measurement preserves both the local and global features of image, meanwhile, it also takes the CRR contribution differences of different plots into consideration. Experimental results show this measurement improves recognition performance significantly.(5) Viewing the difficulty to recognize face in images taken under different conditions, a novel recognition method employing WPD and (2D)2PCA is developed. Firstly, 20 plots are obtained via two-level WPD on the original image. Secondly, the CRRs of these plots are gained by (2D)2PCA and the nearest neighborhood classifier, and‘successful’plots are selected based on these CRRs. Thirdly, the distance between testing image and each database image is calcualted using the proposed distance measurement. Finally, the nearest neighborhood classifier is adopted for recognition on the basis of this distance.(6) The proposed recognition mehod is accomplished by MATLAB 7.0 and images from CMU PIE, Yale or UMIST databases are selected to test the recognition improvement of the proposed method under different illumination, expressions and poses respectively. The performance of (2D)2PCA on the original image is defined as‘standard’method. As the experimental results suggest, the proposed method obtains better performance than‘standard’method under these three conditions. It performs best under different illumination whereas its performance decreases slightly under different expressions and is worst when poses change, and its highest CRRs are 98.795%, 89.796%, 36.047% respectively.(7) Observed from experimental results, the choice of distance metric has a significant effect on face recognition. In general, L1 shows higher CRRs on approximation plots, whilst L2 performs better on detailed plots. Similarly, the filters also show different performances under three different conditions. Therefore, distance metrics and filters should be selected according to these conditions. In the concrete, L1, Daubechies4, and A1, A2, H2, V2, HH2 are recommended to form the proposed method under different illumination, and L1, Haar and A2 are recommended to form the proposed method under different expressions. (8) The proposed method fails to gain satisfactory CRRs under different poses, and the highest record is 36.047%. Thus, it is necessary to seek other methods to extract facial features more efficiently.更多还原