【作者】 谢芳；

【导师】 邹玮；

【作者基本信息】 苏州大学 ， 电子与通信工程（专业学位）， 2013， 硕士

【摘要】 随着社会的发展，各领域对快速有效的自动身份验证的要求日益迫切。生物特征以其自身稳定性和个体差异性成为了身份验证的理想依据。其中，人脸特征具有直接、友好、方便的特点，成为了身份验证最直接自然的手段。自动人脸识别技术就是利用计算机分析人脸图像特征并提取有效信息，以此确认身份的一门技术，涉及计算机视觉、模式识别、图像处理、生理学、心理学以及认知学等诸多学科领域，极富研究的挑战性。人脸识别应用中人脸图像易受实际光照、人脸表情或者姿态的影响，如果特征提取不恰当，很容易造成识别效果的不佳。本文主要针对人脸图像所包含的上述特点，在特征提取与分类角度着手，主要做如下的研究：首先研究分析了目前一些主流的人脸识别算法，其中，小波以其独特的多分辨率分析的特点在信号处理领域得到了广泛的应用。而人脸图像也可作为一种信号输入，通过小波变换实现空域和频域的多尺度分解，得到不同层度的信息，获得较好的分类效果。由此，本文在已有的研究成果上就小波变换的两个关键问题—小波基函数以及小波分解层数的选择进行了深入探讨与研究，通过理论分析和对人脸库的测试实验，验证了小波基函数多样的特征属性对于人脸特征提取的不同表现及其对分类判别带来的影响；以及不同尺度分解对于弱化人脸受光照、姿态与表情影响的不同效果，可望为小波变换在人脸识别的应用中对于小波函数和分解尺度的选择提供有益的参考。其次，研究了基于主成份分析(Principal Component Analysis, PCA)的人脸识别。主成份分析是用于人脸识别的经典特征提取算法，包括一维的主成份分析与二维的主成份分析（广义主成份分析）。本文在研究特征级融合小波与主成份分析的基础上，提出了决策级融合的方法，以期利用融合分类器之间信息互补的特点，提高整体的识别率。算法首先分别采用小波变换与主成份分析提取特征并采用隶属度的方式进行一级模糊分类，其次给予分类器不同的权值，实现加权融合，实验结果表明，选取恰当的权值，可以达到信息互补的目的，提高识别率。最后，研究了采用模糊积分融合分类器的人脸识别算法。模糊积分能够充分考虑各个分类器的重要性以及分类器之间的相互影响和交互作用。用模糊积分的方法将小波变换和二维主成分分析(2-Dimension Principal Component Analysis，2DPCA)进行决策级融合，首先将各个分类器对应各类别的模糊测度与对应每类的模糊隶属度进行模糊积分，选择最大的积分值对应的类为最终的判别结果。其中，模糊测度反应了分类器的特性与交互能力，主要根据样本分类的正确率，描述品质和可分辨度进行综合评价。实验结果表明该算法较单分类器或者一般的加权融合都有着较高的识别率。更多还原

【Abstract】 As the development of the society, automatic verification of identity is required moreand more imperatively in most of the fields. Biometrics is recognized as a reasonable basisfor authentication which is attributed for its self-stability and individual diversity.Therefore, with the metrics of direct, friendly and convenient, auto face recognition madethe most immediate way among all of the verification methods. Automatic face recognitionis a technology which recognizes faces according to face image analysis and featureextraction. There are full of challenges in this study that related to the fields of computervision, pattern recognition, image process, physiology, psychology, cognitive sciences andso on.The present difficulty of the face recognition study is mainly on the poor featureabstraction due to the effect of the light, emotional expression, poses and the like, whichfinally reaches an unwanted results. To deal with problems mentioned above, what themain work have been done are as follows:Firstly, the present popular methods of face recognition are studied, among which,wavelet analysis plays an important role in the signal processing field attributed to thetraits of the multi-resolution analysis. To be treated as one signal, the face can bedecomposed into multi-scales of space and time domains, in which obtains more availableinformation and reaches a better classification result. So, on the base of former researches,the two key problems of wavelet transform including the choices of wavelet anddecomposing scales have been studied here further. Through theoretical analysis andlaboratories testing on face databases, there are some conclusions received such as therelationship between the various characteristics of wavelet function and the manifestationsfor facial feature extraction, the impact of discrimination that has brought out, as well asthe decomposition of different scales for the weakening of the effects on human face bylight, gesture and expression. The research is expected for a useful reference for thechoices of the wavelet function and the decomposing scale in face recognition applications. Secondly, Principal Component Analysis (PCA) has been studied in face recognitionas one of the classic feature extraction algorithm in this fields, which includesone-dimensional and two-dimensional analysis methods. After studying on characteristiclevel fusion, a decision level fusion method has been proposed to improve recognize resultusing information complementary among classifiers. First of all the algorithm, wavelettransform and PCA have been taken to do feature extraction, then, fuzzy classification isdone according to membership approach. Followed by the primitive classifiers, differentweights have been assigned to make a fusion. Results shows that as weights are rightlychosen, it will achieve information complementary and improve the recognize rate.Finally, the algorithm of fuzzy integral fusion applied in face recognition is studied.With the well reflection of the interaction between classifiers, the fuzzy integral hasapplied in many fields and got fine results. Here takes the fuzzy integral to fusion twoclassifiers which classified by the features extracted through wavelet transform and2DPCA. The algorithm is characterized by the integral using fuzzy measure of differentcategories which showed the traits and interactivity to receive the value of all the classes,which chooses the largest integral value as the final result. As one important element in theintegral algorithm, fuzzy measure can be evaluated comprehensively from the therefollowing points as correctness of recognition, describing quality and distinguishabledegree. Experimental results show that the algorithm performs better than a singleclassifier or a formal fusion method using weighs and is applicable in face recognitionsystem.更多还原