【作者】 张博；

【导师】 景晓军；

【作者基本信息】 北京邮电大学 ， 通信与信息系统， 2012， 博士

【摘要】 随着公共和个人安全应用(比如出入境、交通运输安全、网络登录和访问控制等)的发展,身份识别技术已经成为当前研究的热点。其中生物特征识别是利用独特的生理特征和行为特征来进行个人身份认证识别。由于指纹、虹膜、掌纹、人脸、声纹等生物特征具有唯一性、不变性、可靠性和便于采样的优点,生物识别已经成为最有效的个人身份认证之一。人类使用指纹识别身份已有几百年历史,特别是进入19世纪之后,指纹识别技术被应用于科学研究领域,得到了更好的发展。指纹识别技术凭借其很高的实用性和可行性,成为目前应用最普遍并且具有法律效力的生物识别技术。目前,指纹识别中,可获得的指纹往往出现污损、伤疤、断裂等情况。低质量的残缺指纹由于特征丢失过多和非线性形变严重,识别比较困难。但在一些特殊场合和环境下,对残缺指纹的识别却是非常必要的。因此迫切需要对该方面开展研究。本文主要从残缺指纹识别中方向场重建、分割、感兴趣区域提取和匹配四个方面进行研究并提出相关算法。本文主要贡献如下：1、研究并改进残缺指纹方向场重建算法。指纹的脊线方向反映了指纹纹理特征。方向场计算的正确性对后续指纹识别过程中指纹图像增强和匹配有着重要影响。因此对于残缺指纹识别而言,第一步就是对方向场的重建。同时,也要对残缺指纹中的噪声、污损和断裂部分进行去除和恢复。在现有计算指纹方向场算法基础上,结合残缺指纹识别的需要,针对方向场的特点,本研究提出了一种新的描述指纹局部信息和全局信息的描述子。该描述子反映了局部信息的竞争性,全局信息的互补性和冗余性,能够为后续计算提供融合信息。另一方面,提出基于熵理论的指纹方向场算法,将熵理论用于衡量残缺指纹方向信息的不确定性。通过提出方向信息熵,综合设计方向场恢复和重建方案,力求为预处理和后续分类和匹配提供正确有效的信息。2、研究基于数据挖掘技术的机器学习和非统计模式的残缺指纹分割融合算法。指纹图像可分为两部分：(1)前景区域：手指指尖与传感器相接触的部分,该区域包含大量指纹纹理信息,可以为指纹分类和识别提供有用的特征信息；(2)背景区域：包含各类噪声的部分。在残缺指纹中,缺少有效纹理信息并发生严重的非线性形变,因此,有必要对残缺指纹进行分割,在充分保留前景区域正确特征的同时,尽可能的去除背景区域的错误特征。本研究提出了一种利用指纹灰度和局部梯度离散度融合局部指纹梯度信息,并以定义的指纹残缺部分的熵和邻域方向一致性为特征,训练用于指纹分割的支持向量机(SVM)。采用少量训练样本,在复杂的残缺指纹中,得到较为普适的分类器。另一方面,另一方面,针对SVM的缺点,利用非统计的局部二值模式(LBP)纹理特征衡量相邻指纹块之间的竞争性和一致性。最终,结合LBP描述的纹理特征和基于方向熵的SVM的输出结果来做出最优分割。3、研究基于方向熵的残缺指纹感兴趣区域提取算法。指纹的两类最重要的全局特征是中心点和三角点,它们具有平移不变性、旋转不变性、伸缩不变性。通过度量库指纹和查询指纹的全局特征或局部特征之间的相似性就可以判断两副指纹是否来自同一手指了。另一方面,在指纹匹配中,参考点通常被定义为指纹脊线曲率最大的点。通常,中心点被选为指纹的参考点。此外,很多匹配算法中,都是基于对参考点周围的指纹感兴趣区域(ROI)进行匹配。因此,指纹感兴趣区域的确定取决于参考点的精确定位。本研究在现有利用Poincare Index算法的基础上,利用自定义的方向熵,综合提取感兴趣区域。Poincare Index是经典算法,但是需要辅以其他算法排除伪中心点,从而进一步精确提取感兴趣区域。自定义方向熵能够从方向场角度反映指纹区域的变化,能够为提取感兴趣区域提供有效信息。4、研究基于信息融合理论的残缺指纹识别算法。匹配算法是指纹识别的核心。指纹匹配算法可以大致分为基于细节点的匹配算法、基于相关性的匹配算法和混合匹配算法三大类。基于细节点的匹配算法,虽然对细节点提取的精度依赖性很大,但是也是最常用的匹配算法。基于相关性的匹配算法计算量和复杂度比基于细节点的匹配算法要低,但是它对位置、尺度和旋转都比较敏感。混合匹配算法利用细节点和指纹全局特征,将基于细节点匹配和基于全局指纹匹配的算法结合起来得到最终的匹配分数,但是这种混合分数仅仅是简单的加和。然而,残缺指纹信息缺失严重,单一匹配算法针对性较强,难以有效识别复杂的残缺指纹。本研究对指纹识别算法的多子源融合问题进行了深入的讨论与研究,设计基于信息融合理论的指纹识别算法,利用多匹配准则决策层融合,实现残缺指纹有效识别。更多还原

【Abstract】 With the development of the public and individual safety applications (e.g., border crossing, transportation security, network logon, access control, etc.), identification technology becomes the current hot topic for researchers. For automatic identification, biometric recognition refers to the use of distinctive anatomical and behavioral characteristics. Due to its uniqueness, immutability, reliability and convenience for collection, biometrics, such as fingerprint, iris, palm, face, voice, has become one of the most effective personal identity recognition. Fingerprint has been applied to recognition for hundreds of years. Especially since the19th century, it has experienced a better development in scientific research field. Nowadays, fingerprint recognition is the most widely used biometric recognition with legal effects since it is highly feasible and practical.Currently, the acquired fingerprints often have defects in fingerprint recognition such as dirty parts, scars, creases and so on. Due to the large lose of texture and serious nonlinear deformation, it is actually difficult for the low quality incomplete fingerprint to be identified. However, in some special circumstances, incomplete fingerprint recognition is inevitable. Therefore, it is important to do some researches on this project.This thesis mainly determines four key aspects in terms of incomplete fingerprint recognition, mainly orientation field reconstruction, segmentation, Region Of Interest (ROI) extraction and matching. Our major contributions are briefly shown as follows:1. The algorithm of incomplete fingerprint orientation field estimation is improved and proposed.Fingerprint ridge orientation reflects the characteristics of the fingerprint texture. The computational accuracy of orientation field directly affects the results of fingerprint enhancement and matching. So orientation reconstruction is the first step for incomplete fingerprint recognition. Besides the noises, the dirty parts and the creases should have been removed or restored. Based on the existing algorithms, a novel descriptor, including local and global information, is proposed for the characteristics of incomplete fingerprint. According to the properties of competition of the local fingerprint, complementation and redundancy of the global fingerprint, the fusion information is provided for the next step, adaptively. Moreover, the algorithm of incomplete fingerprint orientation field estimation is proposed based on entropy theory. And the uncertainty of incomplete fingerprint is shown as the orientation entropy, with which the fingerprint orientation field of the incomplete area is re-computed and measured. And the fusion orientation field is used for pre-processing and matching.2. The fusion segmentation method is proposed for incomplete fingerprint based on data mining, machine learning and non-statistical pattern.The fingerprint image can be divided into two parts:(1) The foreground that the finger pressed on the sensor, which contains a lot of texture information and provides useful features for fingerprint classification and recognition.(2) The background that contains all kinds of noises. For incomplete fingerprint, fingerprint matching becomes very difficult due to the loss of useful information and serious nonlinear deformation. As a result, it is necessary for incomplete fingerprint segmentation to include the correct features in the foreground and simultaneously remove the wrong features in the background efficiently. The feature vector, based on the gray, the gradient dispersion, orientation entropy and orientation coherence, is defined for segmentation with Support Vector Machines (SVM). With a few training samples, the general classifier is attained in the complex image of incomplete fingerprint. Furthermore, according to the defects of SVM, the texture feature of non-statistical LBP is used to measure the correlation and competition of direction among neighborhoods. Finally, the texture description by LBP and the output of SVM based on the orientation entropy are combined together to get the optimal segmentation.3. The method based on orientation entropy is proposed for incomplete fingerprint to extract the Region of Interest (ROI).The global characteristics, core and delta, are invariant to translation, rotation, expansion and reduction of fingerprints. By measuring the similarity of global or local features in the stored templates and the query fingerprint image, it can be decided whether the two fingerprints are matched or not. Moreover, in matching, a reference point is defined as the point that has the maximum curvature in the most internal ridge. Usually, a core point, the topmost or bottommost point on the innermost recurving ridgeline, is used as such a reference point. And some matching methods are only executed on the ROI centered at the reference point. Therefore, the ROI determination relies on the accurate detection of the reference point. The proposed method is based on orientation entropy and Poincare Index to detect the reference point of fingerprints and extract ROI subsequently. Poincare Index is classic, but it has its defects. In order to extract the ROI. Poincare Index should be complemented by other algorithms to exclude pseudo core. Reflecting changes in the texture with the perspective of the orientation field, the proposed orientation entropy can provide useful information to extract the ROI.4. The fingerprint recognition algorithm based on information fusion is proposed.The matching is the key point in fingerprint recognition. Methods of fingerprint matching can be coarsely categorized into:minutia-based, correlation-based and hybrid-based. The minutia-based algorithms, which rely on the accurate extraction of minutia, are common method of identification. Correlation-based algorithms require less computational complexity than minutia-based methods, but they are vulnerable to variations in position, scale, and rotation. The hybrid-based methods use both minutia and feature information. They combine minutia and feature matching results to generate a final matching score, which is a simple addition. The incomplete fingerprint has serious loss of texture and nonlinear deformation, while the single matching algorithm is hard to identify the complex incomplete fingerprints. As a result, the sub-matching fusion is in-depth discussed and researched in fingerprint recognition. Besides, the fusion matching algorithm has been proposed for the incomplete fingerprint recognition using the fusion decision criterion of different matching algorithms.更多还原