【作者】 刘京；

【导师】 谭铁牛； 汪增福；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2014， 博士

【摘要】 随着社会信息化进程的深入发展,物联网和移动互联网中的光学图像获取装置呈现泛在发展的趋势,同时互联网中的数据规模呈爆炸式增长,为虹膜识别提供了新的数据获取途径。但在此泛在视觉感知环境中获得的虹膜图像存在光源波段、清晰程度、成像场景、用户状态等方面上的显著差异,造成了异质虹膜图像识别的问题。这一问题超越了传统虹膜识别算法的处理能力范围,我们必须寻求新的解决方案。本文从图像重建、特征分析、距离度量和信息融合等多个层次研究异质虹膜图像识别的理论和方法,通过图像处理和模式识别的方法消除成像距离、场景和器件等因素对异质虹膜数据的外在影响,探寻隐藏在复杂多变虹膜数据中稳定不变的身份关联信息,实现稳定、准确的异质虹膜图像识别。本文首先分别针对模糊和低分辨率这两个在异质虹膜识别中最为重要的问题进行讨论,然后提出普适的识别框架。综上,本文主要工作包括：·模糊虹膜图像中丢失的纹理细节是现阶段影响识别性能的主要问题之一。以减少清晰和模糊的异质虹膜图像之间的表观差异为目标,本文提出基于点扩散函数修正的普适虹膜图像去模糊算法。首先,将输入图像按模糊种类划分为离焦和运动模糊,在参数化模型下对两种情况中的点扩散函数进行初始化。然后,在像素级自由度上对点扩散函数进行修正,即对点扩散函数优化、有效区域选择和清晰图像估计三个步骤迭代求解。最终用恢复后的图像进行虹膜识别以获得性能的提升。·虹膜图像作为识别样本,在去模糊过程中应更多考虑与识别相关的内容,而不仅关注于视觉效果。因此,本文提出了层级化的图像先验模型,对不同用途的区域自适应的选择先验学习方法：虹膜区域使用基于特征选择的方法以满足机器感知的要求,而眼周区域根据视觉属性构建先验模型。为了能够有效利用这一模型且进一步灵活融合其它先验知识,本文提出基于隐变量的公式化表达。此方法以识别为导向,恢复出的虹膜图像充分考虑计算机所敏感的纹理细节,提供更有价值的图像增强结果。·实验证明大尺度运动模糊的虹膜图像去模糊后进行识别的结果仍有提升空间。我们通过实验揭露运动模糊影响识别性能的模式及其本质原因之后,提出基于加权匹配模板的比对策略。在第一种模板生成方式中,对观测到的影响模式加以利用,根据每幅图像的模糊角度自适应的设置加权模板。在第二种方法中,利用训练得到运动模糊时虹膜编码中不稳定的区域得到加权模板。这种运动模糊的解决策略更加直接、稳定和高效。·除图像模糊之外,低分辨率是另一个困扰非可控环境中虹膜识别的重要问题。本文基于度量学习将异质虹膜样本映射到特定的度量空间以消除高低分辨率的异质虹膜图像间的差异。在此算法中,我们找到一个变换,将数据集中的异质比对(高、低分辨图像间比对)样本坍塌到对应的同质比对(高分辨率图像间的比对)样本后,再进一步减小类内差异并增大类间差异。然后,学习一个马氏距离使其最大程度的继承理想变换中有用信息。此距离度量能更准确的分别跨分辨率异质虹膜识别中的类内和类间比对,保证更优的识别结果。·以普适异质虹膜识别方法为目的,提出一个编码层的信息映射方法将测试的异质编码映射到对应的注册状态,消除不同状态编码间的异质性后再识别。我们使用改进后的马尔科夫网络对注册状态和测试状态异质编码间的非线性联系进行建模。同时,根据不同编码位所能取到相容度的最大值来衡量其可靠性,并优化得到一个基于统计信息的加权匹配模板。进一步将此方法扩展到多异质源的情况,使得其能够根据不同观测样本选择对应的先验知识。所提编码层方法能在可分性和鲁棒性中找到合理的折中。综上所述,本文以异质虹膜识别为主线,在经典虹膜识别的框架下,按照所处理异质源的不同,提出了一系列异质虹膜识别的解决方案,从而提升系统对于异质样本的容忍度,在针对多源异质虹膜图像的识别算法研究中做出了有益的研究贡献。更多还原

【Abstract】 The Visual Internet of Things and Mobile-Internet provide ubiquitous sensors and processors for acquisition and recognition of iris images. However, in this scenario, it is inevitable to capture a large number of heterogeneous iris images with quality and condition variations. In this thesis, the theory and method for robust recognition of heterogeneous iris images are systematically studied. First, blur and low resolution which are of great importance in heterogeneous iris recognition are discussed respectively, and then a general framework is proposed. The main contributions are as follows:· Blurred iris images without texture details degrade recognition performance. A novel deblurring algorithm based on point spread function (PSF) refine-ment is proposed to alleviate the visual difference between clear and blurred iris images. First, the input images are classified to be defocus or motion blurred, and PSF is initialized based on parametric models. Second, PSF is refined in pixel level by iteratively optimization of PSF, support regions and the latent image. Finally, the deblurred images are applied for recognition, which improves recognition accuracy.· As recognition samples, iris images should be deblurred with emphasis on recognition items rather than visual effects. Hence, a hierarchical model is proposed to adaptively assign prior learning methods to regions with d-ifferent usages, which enables both visual and machine perceptions in iris images. The latent variable based formulation incorporates the hierarchical model and other task-specific terms, which guarantees its flexibility.· The recognition performance of motion blurred images after restoration is shown can be further improved. Experiments are conducted to reveal the relation between distortion and performance. A mask-based matching s-trategy is proposed. In the first type of mask generation, this observation is used to generate the mask. In the second type, based on training samples, the reliability of each code bit is estimated and used for mask generation.· Apart from image blur, low resolution (LR) is another important problem in heterogeneous iris recognition. A metric learning based algorithm is pro-posed as the solution in metric space. In the method, an ideal mapping is defined to collapse heterogeneous (High-resolution vs. LR) and homoge-neous (HR vs. HR) samples, and keep identically labeled samples close while separating samples in different classes. The target Mahalanobis distance is learnt to capture the information in ideal mapping as much as possible. Based on this learnt metric, it is easier to sperate intra-and inter-class samples, which enhances recognition performance.· Aiming at a general solution, we propose a code-level information mapping algorithm. It can map the probe-state iris codes into the corresponding reg-istered states and then uses the mapped codes for recognition. Code-level operations are sandwiched between feature and score levels. The modified Markov model is employed to model the nonlinear relationship between het-erogeneous codes. Meanwhile, the maximum compatibility value is used to measure the bit reliability and optimized to form a matching mask. This algorithm is also extend into the situation of multi-source heterogeneity.In this thesis, we focus on heterogeneous iris recognition and follow the typical recognition procedure. According to different sources of heterogeneity, solutions are proposed to make the recognition systems more tolerant to heterogeneous samples, which contributes to the research area of heterogeneous iris recognition.更多还原