【作者】 韩民；

【导师】 彭玉华； 王洪君；

【作者基本信息】 山东大学 ， 通信与信息系统， 2010， 博士

【摘要】 随着当今信息化社会的发展,个人身份鉴别的重要性正日益显现,上至国家安全、下至百姓生活,无不对其提出越来越高的要求。对于我国这样一个人口大国,身份鉴别有着尤其广泛的应用前景和重要的战略意义。生物特征认证,作为上个世纪末期才开始蓬勃发展的身份鉴别技术,利用人本身所拥有的生理特征和行为特征来判别个人身份,必将逐步取代传统的身份鉴别方法,在社会生活中占据越来越重要的地位。虹膜识别技术是近年来逐步兴起的生物特征识别技术,以其高精确度、非接触式采集、易于使用等优点日益引起学术界和企业界的关注,被广泛认为是未来最具有发展前途的生物特征识别技术。多分辨率分析是现代图像处理一种很重要的工具,能够对图像进行多尺度的分解和表示,以寻找单一分辨率下无法发现的特性,更有效的提取图像特征,获取图像信息。本文以基于人体虹膜的生物特征识别技术研究为题,在深入研究多分辨率分析理论的最新进展和虹膜识别的主要特点的基础上,重点围绕多尺度下虹膜图像特征的有效提取以及与其相适应的虹膜图像预处理和特征匹配算法,展开一系列有益的探索。本文首先简要综述了生物特征识别的基本原理以及目前常用的生物特征识别技术,重点对虹膜识别技术的主要特点、发展历史、关键组成部分及其研究现状进行了详细的分析和总结；接下来,在对现有传统虹膜预处理方法进行深入研究的基础上,提出一种更有利于二维虹膜图像纹理信息特征提取的虹膜图像归一化方法；然后基于对多尺度几何分析理论的研究,采用一种经典的多尺度几何分析方法——Contourlet变换提取虹膜图像的多尺度多方向性信息作为虹膜特征矢量,应用支持向量机对虹膜进行分类和识别；进而本文针对Contourlet变换中实现尺度分解的拉普拉斯金字塔存在的一些不足,在对目前常用多分辨率分析算法进行比较分析后,采用一种性能更加优越的圆周对称滤波器应用于虹膜图像的多分辨率分解,以获取更加稳定而准确的虹膜特征,获得更好的识别性能。本文还将近年来逐渐受到研究者关注的自适应非平稳信号分析方法——经验模态分解应用于虹膜识别,通过实验分析捕捉到最适合虹膜分类的频带进行虹膜特征编码,取得了理想的识别性能。本文的主要创新之处在于：1、提出一种非极坐标展开的虹膜图像归一化方案传统的虹膜图像归一化方案都是在极坐标下,采用映射的方式将类似环形的虹膜图像沿角度方向展开为固定尺寸的矩形,以实现虹膜图像的尺度不变性和平移不变性。这种方法简易有效,被以往所有的虹膜识别系统所采用。然而。由于虹膜内外圆半径具有较大差异,这一方案在展开的过程中势必进行大量的插值或抽取,改变了原始虹膜图像纹理几何结构和方向信息,而且由于要避开眼睑与睫毛遮挡区域,实际使用的虹膜有效区域是一个狭窄的长方形图像,这对于以往采用局部信息或将二维虹膜图像转化为一维信号进行特征提取的算法影响可能不大,当采用虹膜全局信息或纹理方向信息作为特征的算法,难免会产生不利的影响。为此本文提出一种不需要将其极坐标展开为长条矩形的归一化方法,保留了虹膜图像纹理的原始几何结构和方向信息,利用周边原始虹膜图像反转映射的策略填充内圆中空区域,以免引入新的边界。得到的图像在尺寸上更加均衡,并且可以根据虹膜图像采集的质量选取有效区域,灵活的避开眼睑与睫毛的遮挡,为多尺度几何分析或其他特征提取算法提供了更多的选择。2、提出一种基于Contourlet变换和支持向量机的虹膜识别系统多尺度几何分析方法和理论是近年来逐渐兴起的图像处理方法,其在对二维图像的稀疏表示上所展示出的优越性能倍受人们称赞。本文采用一种经典的多尺度几何分析算法——Contourlet变换作为工具提取虹膜图像全局纹理信息,以经过Contourlet变换分解后的多尺度多方向子带的归一化能量作为特征矢量,应用支持向量机进行分类识别,通过实验,本文遍历性的尝试了不同的尺度分解级数和各尺度下不同的分解方向数,列出各种方案下的识别率,为多尺度几何分析在虹膜识别中的进一步应用研究提供了有益的探索和指导。3、提出一种改进的圆周对称滤波器组应用于虹膜特征提取的方法多分辨率分析理论经过二十多年的发展,已成为图像处理不可或缺的重要工具,实现图像多分辨率分析的简单有效的结构就是一系列分辨率逐步降低的金字塔形图像集合,对图像不同分辨率下的分析非常有利于模式识别。本文在简要分析比较目前常用金字塔分解算法的基础上,采用一种改进的圆周对称滤波器组用于虹膜图像的多分辨率分解,不仅可以从理论上消除频谱混叠现象,而且具有平移和旋转不变性,使虹膜图像各尺度上的分量更加准确和稳定；并结合方向滤波器组对虹膜图像进行各尺度下的方向分解,获取虹膜多尺度多方向性信息进行识别,实验结果表明,这种方法可以获得更好的识别效果。4、提出一种基于经验模态分解的虹膜识别算法经验模态分解是一种新颖的信号分析方法,它可以自适应的将非平稳、非线性信号分解成一组稳态、线性的固有模态函数分量,这些固有模态函数分量相互独立,突出了信号的局部特征,对应着不同的频带信息,从而实现信号的多分辨率分解。本文利用这一特性,将虹膜图像分解为一系列按频率由高到低排列的固有模态函数分量,以互信息测度为初步评判依据,并进一步通过大量实验分析筛选出有利于虹膜识别的固有模态函数分量及其组合,对其进行有效的特征编码,并使用运算量较小的汉明距离匹配算法进行分类识别。这一方法不仅可以简化虹膜图像预处理过程,而且可以达到很高的识别性能。本文的工作是为了丰富现有虹膜识别算法而进行的一些有益的探索和尝试,实验结果表明,所提出的虹膜识别算法基本满足实际应用的要求,在此基础上开展进一步的研究并加以完善,可以开发出我国具有自主知识产权的虹膜识别系统。更多还原

【Abstract】 With the development of information technology, the importance of personal authentication is becoming increasingly apparent. Higher and higher requirements are imposed on it from national security to everyone’s daily life. For our country with so large a population, identity authentication has broad application prospects and strategic importance. Biometric identification, as an authentication technology with flourishing development at the end of the last century, makes use of a person’s physiological and behavior characteristics to determine personal identity. It plays an increasingly important role in our social life and will gradually take the place of the traditional identification methods. Iris recognition is a biometric recognition technology emerging in recent years. It has attracted the attentions of academic and business community for its advantages such as high accuracy, non-contact collection, easy to use etc. It is widely considered as the most promising biometric recognition technology in the future.Multi-resolution analysis is a very important tool in modern digital image processing. It can represent an image in a multi-scale manner and can be used to find the image characteristics and extract image features unavailable in the pixel domain. With Iris recognition technology as the topic and based on the deeply study on the multi-resolution theory and the main features of iris recognition, this thesis researches on the efficient feature extraction algorithms of Iris images under multi-resolution framework and corresponding preprocessing and feature matching algorithms. Firstly, the principle of biometric recognition and the commonly used biometric recognition methods are introduced briefly, with emphasis on the main features, history, the key components and the current research status of iris recognition technology. Then, after investigating the existing preprocessing methods in depth, a normalization method for the Iris image is proposed which is more suitable for extracting the texture features from the 2D Iris image. Subsequently, based on the Multi-scale Geometric Analysis (MGA) theory, the classic contourlet transform is adopted to extract the multi-scale and multi-directional information as the feature vector, and support vector machine is used to classify these features for recognition. Next, for the deficiency of the Laplacian pyramid in contourlet transform, after comparing the commonly used multi-resolution analysis methods, a superior circular symmetric filter is applied to decompose the Iris image in order to get more stable and accurate Iris features and thus better the recognition performance. In addition, an adaptive non-stationary signal analysis method-empirical mode decomposition is used for Iris recognition, the frequency bands most suitable for recognition are identified through experiments and ideal recognition performance is achieved.The main innovations of this thesis include:1. A normalization scheme of the iris image is proposed with non-polar expansionIn traditional methods, the Iris image is normalized under polar coordinate, the annular-like Iris image is mapped into a rectangular region of fixed size along the angular direction, in this way, the scale and translation invariance can be achieved. This method is simple and effective, so it is widely used in all previous Iris recognition systems. However, because of the difference between the radii of inner and outer circles, a large number of interpolation and decimation operations must be taken during the expansion process, the texture structure and directional information of the Iris image are inevitably altered. And what’s more, in order to avoid the regions with eyelids and eyelashes occlusion, the available region is just a narrow rectangular image, this may take little effect on the feature extraction methods of using local information or transforming the Iris image into a 1D signal. But surely, it will make a negative impact on the algorithms with global information or directional information as the features. To solve this problem, a normalization method is proposed in which there is no need to expand the polar coordinate into a rectangular. This method can preserve the original geometric structure and directional information of the Iris texture, and a reverse mapping strategy is used to fill in the region in the inner circle avoiding introducing new boundaries. The obtained normalized image is more balanced in size, and the effective region for Iris recognition can be selected flexibly according to the acquisition quality, which provides more choices for the feature extraction methods using multi-scale geometric analysis and other methods.2. An Iris recognition system based on contourlet transform and SVM is proposedMulti-scale Geometric Analysis methods have superior performance in sparsely representing the 2D image. In this thesis, contourlet transform is used to extract the global texture information of an Iris image, the normalized energies of the multi-scale and multi-directional subbands act as the feature vector, and the SVM is adopted for recognition. Experiments have been designed to try every possible combination of the decomposition level and number of directions used in an exhaustive manner, and the recognition rates under all schemes are listed, which can provide an instruction for further research on the Iris recognition systems based on MGA.3. An Iris feature extraction method based on an improved circular symmetric multi-resolution decomposition scheme is proposedThe Multi-resolution analysis theory has become the indispensable tool in image processing community. The analysis process under different resolutions has advantageous for pattern recognition. Based on the comparisons of the classic pyramid decomposition algorithms, an improved circular symmetric filter bank is used for Iris image decomposition. It avoids the frequency aliasing effect and owns translation and rotation invariance, which makes the Iris information in different scales more accurate and stable. Then the directional filter bank is adopted to get the directional information for recognition. Experimental results show that the proposed scheme can achieve better recognition rates.4. An Iris recognition approach based on empirical mode decomposition is proposedThe empirical mode decomposition(EMD) can adaptively decompose a non-stationary, nonlinear signal into a series of stationary and linear intrinsic mode functions(IMF). These IMFs are independent of each other and reflect the local characteristics of a signal. EMD can be regarded as a multi-resolution decomposition method with each IMF corresponding to a different frequency band. In the proposed method, the Iris image is decomposed by EMD into a few IMFs from high to low frequencies, then the IMFs or their combinations most suitable for recognition are determined with mutual information criterion as a primitive evaluation and large amount of experimental analysis. The selected IMF or IMFs are coded effectively and Hamming distance is adopted as the matching algorithm for recognition. Experiment results show that this method can not only simplify the preprocessing procedure, but can achieve surprisingly high recognition performance.The work in this thesis can be regarded as some useful exploration and attempt in order to enrich the existing Iris recognition algorithms. Experimental results show that the proposed iris recognition algorithms can basically meet the requirements of practical applications. Based on this work, further deep researches can be conducted to improve the algorithms’ performance, and an iris recognition system with independent intellectual property of China is expected to be developed.further research can be carried out based on it.更多还原