【作者】 潘力立；

【导师】 解梅；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2012， 博士

【摘要】 虹膜识别作为生物特征识别的一个重要分支，是指利用人眼虹膜区域进行身份鉴别的技术。虹膜特征与其它生物特征相比具有高稳定性、高可靠性和高防伪性等优点。该技术从提出到现在已有十多年的时间，但是由于应用环境的复杂性，仍然存在许多需要改进之处。本文正是针对现有虹膜识别算法在虹膜图像模糊检测和虹膜图像定位两部分的不足之处，提出了新的改进性方法，并取得了较好的实验结果。本文首先介绍了生物特征识别技术和虹膜识别技术的基本概念，以及虹膜识别系统的主要组成部分，即：虹膜图像质量评估、虹膜定位、虹膜图像归一化、虹膜图像遮挡掩膜估计、虹膜图像特征提取、匹配和虹膜图像粗分类。同时介绍了每一部分相关算法的研究现状和发展动态，及存在的主要问题。针对这些主要问题，本文进行了以下五个方面的研究工作：1.提出了基于多核函数学习的模糊虹膜图像检测模型。虹膜图像模糊检测是自动虹膜识别系统中图像质量评估最重要的一个步骤，该问题作为非参考图像模糊检测问题一直未被很好解决。本文在分析模糊虹膜图像频谱和倒谱特征的基础上，提出了两种新的基于频谱能量密度分布和奇异倒谱直方图的模糊特征。提出奇异倒谱直方图分布作为模糊特征也是本文一个重要的贡献之一。该特征是与图像内容独立的具有很强判别性的模糊特征。另外，本文提出采用多核函数学习的方法，融合两种特征进行虹膜图像的模糊检测。实验结果证明了提出特征和多核函数学习的有效性。2.提出了基于概率成对投票的虹膜定位模型。虹膜图像定位是虹膜识别系统中最困难且最重要的问题，它直接关系到识别特征的有效性和算法的实时性。眼睑、睫毛和光斑等的干扰以及拍摄时头部的转动加剧了该问题的复杂性。本文提出了一种基于概率成对投票的圆定位方法。该方法能够有效克服被检测物体的微小形变，不易受到遮挡、噪声的干扰。同时由于在概率框架下在连续参数空间中进行投票，因此可以通过模态检测的方法找到最佳假设圆参数，有效地降低了已有圆定位方法的算法复杂度。本文将该方法应用于虹膜定位问题，取得了很好的效果。3.提出了基于局部判别模型和哈夫聚类的虹膜定位模型。多数已有虹膜定位方法利用虹膜边缘的灰度梯度信息进行虹膜边缘的检测或判断，然而虹膜的内外边缘除了灰度梯度值特征之外，还有许多判别性更强的局部特征。本文提出利用边界点附近局部区域的图像特征进行虹膜和非虹膜边界点的判定。同时利用哈夫聚类的基本思想，通过边界点在图像中分布位置的整体约束，进一步准确地区分虹膜和非虹膜边界点。最后，利用光滑样条函数拟合虹膜边界点，得到虹膜边缘的精确位置。实验证明局部判别模型的应用有效地提高了原有虹膜定位方法的定位准确度。4.提出了基于小波多分辨率分析和M估计的虹膜定位模型。小波多分辨率分析是图像分析的有效工具。本文通过分析瞳孔边缘、虹膜边缘以及眼睑、睫毛边缘特征存在的尺度范围，提出在虹膜内外边缘特征存在的尺度范围内进行虹膜边界点提取，有效地减少了其它边界的干扰。同时，提出采用基于M估计的椭圆拟合方法拟合边界点，进一步排除了非虹膜边界点对定位结果的影响。5.提出了基于相位一致性分析和最小修剪方差拟合的虹膜定位模型。相位一致性分析是进行图像特征检测和提取的一种有效方法。本文通过检测相位一致性的局部最大值提取虹膜边界点。该边界点提取方法不易受图像光照变化、对比度变化的影响，同时通过选择滤波器的尺度范围避免了眼睑、睫毛等高频图像特征的干扰。此外，利用最小修剪方差拟合边界点也有效避免了噪声和外点对于拟合结果的干扰。最后，本文总结了上述各种方法的优点与不足，并提出了今后工作的主要研究方向。更多还原

【Abstract】 Iris recognition, as an important branch of biometric recognition, recognizes person’sidentity according to the textures of iris region in our eyes. Compared with otherbiometric identification technologies, iris recognition is much more stable, secure andanti-counterfeiting. In the past more than ten years after this technology was firstlyproposed, many related methods have been proposed for iris recognition. However, theexisting iris recognition methods also have some drawbancks, especially in the irisimage quality evaluation part and iris localization part. Aiming at solving the difficultproblems in these two parts, this thesis proposes a novel improved iris image blurdetection method and four effective iris localization algorithms.At the beginning of this thesis, we introduce the concepts of biometric recognitionand iris recognition. Moreover, we introduce the major parts of iris recongtion system,namely, iris image quality evaluation, iris localization, iris image normalization, irisocclusion mask estimation, feature extraction and matching as well as iris image coarseclassification. In the description for each part, we also introduce the related existingmethods and their advantages and disadvantages. To solve the problems in irisrecognition and improve the performance of exisiting methods, we have done thefollowing works:1. Propose the blurred iris image detection model with multiple kernel learning.Blurred iris image detection is an especially important problem for automatic irisrecognition system. However, it is a non-reference image quality evaluation problemand is hard to find discriminative features for blur detection. In this thesis, we analyzethe characteristics of the frequency spectrum and cepstrum of blurred iris image andpropose two new discriminative blur features, namely: Spectral Ennergy DensityDistribution and Singular Cepstrum Histogram. Defining blur feature in cepstrum isalso one of the major contributions of this thesis. To merge the two proposed blurfeatures, we employ multiple kernel learning theory to construct a merging kernel whichis a linear combinition of two kernels. The experiment results demonstrate the improvedperformance of our method.2. Propose a new iris localization model based on probabilistic pairwise voting. Iris localization is an especially hard and important task in iris recognition. As withother biometric recognition problems, correct localization or registration is key toaccurate iris recognition since it allows “apples-to-apples” comparisons. However, dueto the occlusion from eyelash, eyelid and spectral reflections, as well as the off-axiscapturing, iris localization is a very challenging problem. This thesis proposes a newcircular object detection method which can be used for iris localization. The proposedmethod is robust to small object shape deformations, noise and occlusions. The wholevoting model is formulated in continuous parameter space and the optimal parametercan be detected though using mode-finding mehods. This stratege reduces thecomputational cost of existing voting-based circle detection method. When applying foriris localization, the proposed method performs well.3. Propose a new iris localization model based on local experts of edge point andHough Clustring.For most existing iris localization methods, they usually utilize the gradientinformation of edge points to detect iris inner or outer edge. However, except thegradient value, there are also many other kind of local image patch features which canbe used for iris and non-iris edge point discrimination. In this thesis, we propose toextract the local image feature of every edge point and train two local experts todesriminate iris and non-iris inner or outer edge points. Meanwhile, coupled with thespatial distribution of these edge points, we can slectect the correct iris edge pointsthrough Hough Clustering. Finally, smooth spline fitting is adopted to fit these slectededge points to get the accurate edge of iris. Related experiment results show theimproved performance of this method.4. Propose a novel iris localization method based on Multi-resolution analysis andM-estimation.Multi-resolution analysis is an efficient image analysis theory. In this thesis, wesuppose the edges of pupil, iris, eyelash and eyelid are image features exsisting atdifferent scales. Then, we propose to extract iris edge points at appreciate scales toavoid the disturbing of eyelash and eyelid. Meanwhile, we propose to employM-estimation to fit these extracted edge points.5. Propose a new iris localization method based on phase congruency analysis andtrimmed least square fitting. Phase congruency analysis is an efficient method for image feature localization andextraction. This thesis proposes to detect the local maximum of phase congruency atapprociate scale range and choose these points with local maximal phase congruency aspupillary and iridial edge points. Detecting the edge points of iris based on phasecongruency analysis is robust to illumination change and contrast variations. Moreover,we propose to employ trimmed least square method to fit these edge points robustly.At the end of this thesis, we conclude the advantages and disadvantages of theproposed methods, then introduce the future work.更多还原