【作者】 陈勇；

【导师】 冯浩；

【作者基本信息】 浙江工业大学 ， 控制理论与控制工程， 2010， 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,人脸识别技术的研究具有十分重要的理论价值和应用价值。经过了40多年的研究,人脸识别技术已经取得了重要的发展,涌现出了许多新技术、新方法,在受控环境下,许多算法已能取得令人满意的结果。但是,在光照、姿态变化较大的情况下要实现一个高识别率、高鲁棒性的人脸识别系统仍然是一个极具挑战性的课题。本文对小波变换在人脸识别中的应用进行了深入的思考,在分析当前各种基于小波变换的人脸识别算法的基础上,提出并深入研究了基于样条二进小波的人脸识别问题,论文的主要工作和贡献如下:1、对基于小波变换的人脸识别研究进行了综述小波变换按其连续性不同可以分为连续小波变换、离散小波变换和二进小波变换。根据小波变换类型的不同,将常见的基于小波变换的人脸识别算法分为二大类:(1)基于离散的正交、双正交小波变换的人脸识别方法,简称为离散小波方法。(2)基于连续的Gabor小波变换的人脸识别方法,简称为Gabor小波方法。并从小波在人脸识别中发挥的作用来分析,对这两类方法的特点进行了总结和评价。小波在离散小波方法中主要起到了降维和平滑去噪的作用,因此算法复杂度往往较低。但由于正交、双正交小波提取人脸细节特征的能力并不是很强,使得算法的识别性能还有待提高。而在Gabor小波方法中,Gabor小波具有很强的人脸纹理特征提取能力,因此算法的识别率较高,但由于Gabor小波变换的冗余性很大,造成算法的复杂度较高。2、在理论上合作研究论证了基于二进小波的Mallat分解算法的可行性和有效性首先,介绍了一种新的二维二进小波变换,称为二维平稳二进小波变换,它与正交小波的二维平稳小波变换类似,由一个低频分量和水平、垂直、对角线三个方向的高频分量所定义,这是不同于由Mallat提出的传统三分量形式的二进小波变换。实际上,二维平稳二进小波变换是正交、双正交小波的二维平稳小波变换的扩展;然后,推导并给出了二维平稳二进小波变换的快速分解和重构算法。接着,进一步引入了ε抽样的离散二进小波变换的概念,描述了它与二维平稳二进小波变换之间的关系;最后,讨论了基于二进小波的Mallat分解算法,它是ε抽样的二进离散小波变换的一个特例,此时ε是由全0所组成的一个二进序列,从而从理论上论证了基于二进小波的Mallat分解算法应用的可行性。另外,从理论上分析可知:基于二进小波的Mallat分解与基于正交、双正交小波的Mallat分解具有相同的计算量,但前者是二维平稳二进小波变换系数的一个子集,因此,比后者能够更好地提取图像的边缘特征信息,即在理论上论证了基于二进小波的Mallat分解算法的有效性。最后,与基于正交、双正交小波的Mallat分解算法的应用进行了直观的效果对比,结果显示:基于二进小波的Mallat分解算法能够产生具有非常好边缘效果的细节子图。总之,基于二进小波的Mallat分解算法的提出将进一步拓展二进小波的应用领域,使二进小波有望在模式识别、图像处理等领域取得好的应用,这也是本文提出的所有算法的理论基础。3、较详细地研究了样条二进小波变换在人脸识别中的应用提出了一类新的基于样条二进小波变换的人脸识别方法,简称为样条二进小波方法。在新方法中,小波既具有降维和平滑去噪作用,又有很强的人脸边缘特征提取能力;然后,提出了一种具体应用样条二进小波的人脸识别算法,并以此算法为例详细说明和实验了样条二进小波在人脸识别算法中的应用特点,大量的实验得出了一些有意义的结论:(1)基于样条二进小波分解得到的细节子图包含很多的特征信息,单独应用这些细节子图也具有很高的识别率。相比之下,基于正交、双正交小波分解得到的细节子图包含的特征信息要少的多。(2)样条二进小波分解产生的细节子图所包含的特征信息具有很强的互补性,对多个细节子图的有效融合可以较大地提高算法的识别率。相比之下,正交、双正交小波分解得到的细节子图之间的互补性要差的多。(3)在众多样条二进小波中,其中的正交的样条二进小波和一种非正交的零反对称样条二进小波具有最好的应用效果。总之,样条二进小波应用于人脸识别相比正交小波、双正交小波以及Gabor小波具有许多好的特性,算法在充分利用这些特性后取得了很好的实验结果。因此,基于样条二进小波的人脸识别方法是一个值得进一步深入研究的方向。4、研究了基于样条二进小波的多方向人脸边缘特征提取问题,提出了一种基于多方向人脸边缘特征和二维线性判别分析的人脸识别算法提出了基于样条二进小波变换的多方向边缘细节子图提取方法。通过对人脸图像进行旋转45度、二进小波变换、反向旋转45度以及一定的剪切等一系列简单的操作,可以得到原始人脸图像的45度方向、135度方向以及十字线方向的边缘细节子图。从直观的视觉观察和实验比较证明了该方法是简单有效的。然后提出了一种基于多方向人脸边缘特征结合二维线性判别分析的人脸识别算法,实验证明该算法是有效的。5、提出了两个基于二进小波变换和多分类器融合的人脸识别算法研究了基于样条二进小波变换和多分类器融合的人脸识别算法,针对人脸识别中的表情识别和光照人脸识别问题,提出了两个具体的人脸识别算法。(1)对人脸表情识别进行了研究,提出了一种基于样条二进小波和多分类器融合的人脸表情识别算法。算法在JAFFE人脸库上实现了100%的人脸识别率和85.34%的表情识别率。(2)对人脸光照识别进行了研究,提出了一种结合对数域DCT光照补偿和多分类器融合的光照人脸识别算法,该算法结合了对数域DCT光照补偿、样条二进小波分解、二维线性判别分析以及多分类器融合等多种对光照变化鲁棒的方法。算法在CAS-PEAL人脸库光照子集上实现了最高83.91%的识别率,而在YaleB人脸库上实现了100%的识别率。更多还原

【Abstract】 Face recognition is a biometrie identification technology with best development potential, and researching on face recognition technology has great theoretical and practical values. After more than forty years of research, considerable progress has been made on the problems of face recognition, especially under stable conditions such as small variations in lighting, facial expression and poses. All face recognition algorithms, however, witness a performance drop whenever face appearances are subject to variations by factors such as occlusion, illumination, expression, pose, accessories and aging. So, it’s still a long way to reach a satisfying performance.In this paper, the face recognition algorithms based on wavelet transform are discussed and analyzed. On this basis, a new face recognition approach using spline dyadic wavelet transform is proposed and studied.The principal research work and novelties are listed as follows:1.Provide a detailed survey of face recognition based on wavelet transform.Wavelet transform can be classified into three types, namely continuous wavelet transform, discrete wavelet transform and dyadic wavelet transform. In the last decade, a lot of face recognition algorithms based on wavelet transform have been developed. However, based on the difference of wavelet type used, these face recognition algorithms based on wavelet transform can be classified into two major categories.(1)Face recognition method based on discrete orthogonal or biorthogonal wavelet transform, called Discrete Wavelet Method(DWM) for short.(2)Face recognition method based on continuous Gabor wavelet transform, called Gabor Wavelet Method(GWM) for short. Then, their characteristics of two types of methods are analyzed and summarized from the viewpoint of wavelet application in algorithms. Wavelet application can play the role of dimension reduction and denoising smooth in DWM. However, orthogonal or biorthogonal wavelets are not good at extracting facial edge detail features, which results in the fact that the recognition rate of these algorithms in DWM need to be improved further. On the other hand, Gabor wavelets have good ability to extract facial texture features, which can usually bring good recognition performance. However, it is known to us that GWM have high computation complexity.2. Research and prove jointly the feasibility and effectivity of Mallat decomposition algorithm based on dyadic wavelet.At first, two-dimensional stationary dyadic wavelet transform (2D-SDWT) is introduced, it is defined by approximation coefficients, detail coefficients in horizontal, vertical and diagonal directions, which is essentially the extension of two-dimensional stationary wavelet for orthogonal or biorthogonal wavelet filters. Then, the fast algorithm of 2D-SDWT is given. Next,ε-decimated dyadic discrete wavelet transform(ε-DDDWT) and its relation with 2D-SDWT is given, whereεis a sequence of 0’s and 1’. At last, Mallat decomposition algorithm based on dyadic wavelet is proposed as a special case ofε-DDDWT. The experimental comparison with Mallat decomposition algorithm based on orthogonal or biorthogonal wavelet shows that Mallat decomposition algorithm based on dyadic wavelet has better edge detection effects. The property will further extend the application field of dyadic wavelet. It has great potential in the fields of pattern recognition, image processing and so on. It is also theoretical basis of all algorithms proposed in the subsequent several sections.3. Research the application of spline dyadic wavelet in area of face recognition in detail.present a new type of face recognition method based on spline dyadic wavelet, which is called Spline Dyadic Wavelet Method (SDWM).Dyadic wavelet in SDWM can not only play the role of dimension reduction and denoising smooth, but also extract good facial edge detail features. Next, a concrete face recognition algorithm SDWT-FFT-PCA are introduced, and then take the algorithm for example, analyze application features of spline dyadic wavelet in SDWM. At last, several meaningful conclusions according to some experiments can be given as follows:(1)Detail subbands in SDWM include a lot of feature information, which is revealed by the fact that the algorithm based on single detail subband has also high recognition rate. By contrast, detail subbands in DWM have much less feature information.(2)Different detail subbands in SDWM are complementary each other for face recognition, so the recognition accuracy can be improved greatly after the subbands are fused by an effective fusion strategy. In comparison, it is more difficult for DWM to improve recognition rate by making use of the complementarity of different detail subbands.(3)In many spline dyadic wavelets, the orthogonal spline dyadic wavelet and a nonorthogonal spline dyadic wavelet have the best performance.4.Propose a face recognition algorithm based on multidirectional facial detail features and two-dimensional linear discriminant analysis.A multidirectional detail subbands extraction method based on spline dyadic wavelet is proposed. The method can produce three detail subbands in 45 degree,135 degree and cross direction by a series of operators such as rotating original image 45 degree,dyadic wavelet transform,rotating image 45 degree in negative direction,crop operation and so on. Then , a face recognition algorithm RSDWT-FFT-2DLDA based on multidirectional facial detail features and two-dimensional linear discriminant analysis (2DLDA) is proposed. The experimental results shows that detail subbands extraction method and RSDWT-FFT-2DLDA are effective.5.Propose two face recognition algorithms based on spline dyadic wavelet and fusion of multi-classifiers.Face recognition algorithms based on spline dyadic wavelet and fusion of multi-classifiers are studied. Aiming at facial expression recognition and face recognition under the condition of illumination perturbations, two concrete algorithms are proposed respectively.(1)facial expression recognition is researched, a face recognition algorithm based on multidirectional facial edge detail features and fusion of multi-classifiers is proposed. The algorithm reach 100% face recognition rate and 85.34% expression recognition rate in JAFFE.(2)face recognition under the condition of illumination perturbations are studied. A face recognition algorithm combining illumination compensation with fusion of multi-classifiers is proposed. The algorithm combines discrete cosine transform (DCT) in the logarithm domain, spline dyadic wavelet, 2DLDA, fusion of multi-classifiers respectively. The proposed algorithm are tested on CAS-PEAL and YaleB face databases, and achieve 83.91% and 100% recognition rate respectively.更多还原