【作者】 何飞；

【导师】 刘元宁；

【作者基本信息】 吉林大学 ， 计算机科学与技术， 2015， 博士

【摘要】 在虹膜特征提取和匹配领域的研究中,Gabor滤波器是应用最广泛且实验证明其滤波响应与人类视觉区响应波形一致,被视为最具仿生意义的图形滤波器。本文主要从以下几个着手点研究基于Gabor滤波器的虹膜特征表达和匹配。1.提出了一种自适应Gabor滤波器选取方法。这种方法结合PSO和BPSO优化原则进行Gabor滤波器参数寻优,利用注册图像来训练一组合理的Gabor变换核。这样的Gabor滤波器选取过程完全由训练数据驱动,无需先验知识进行预先选取和人为设定,更能准确导向虹膜纹理信息量最集中的滤波频带进行虹膜特征提取。实验证明本方法可针对不同的采集设备、不同采集环境和不同数据集,优化得到更适合虹膜特征提取的Gabor滤波器,且相应的Gabor滤波器可比经验设定的Gabor滤波器提取更有区分度的特征描述。2.提出对Gabor滤波的幅值和相位进行改进的局部化特征表示,生成虹膜纹理的多Gabor特征描述。本文提出的改进的局部化特征生成方法,可避免直接对图像进行栅格化造成的边缘效应。局部化的幅值特征描述虹膜纹理的能量细节,刻画了纹理的疏密强弱,具有旋转平移不变性；同时,局部化后的相位特征描述虹膜纹理的方向特性,反映纹理的分布趋势,它具有对光照变化不敏感的特点。3.提出一种对虹膜多种特征进行融合的方法。对多特征用相应距离函数求得的相似度进行融合,求取多种特征共同作用下的融合相似度来判定虹膜的分类结果。在求取多特征融合相似度时,我们训练基于SVR模型的估计函数。通过该估计函数,即可将各个特征形成的相似度度量映射到融合相似度,并据此进行虹膜匹配判定。实验证明SVR融合技术充分利用特征的可区分性以提高识别的准确性和稳定性,比其他的虹膜单一特征的识别系统效果更优。4.提出了一种基于SIFT关键点的筛选策略,根据去除相邻像素中能量较低的冗余SIFT关键点,提高SIFT关键点的匹配准确度。同时将Gabor变换域下的局部幅值特征、局部相位特征和该几何域下的SIFT关键点特征,用基于SVR融合方法进行特征融合分类,进一步提高了虹膜识别的鲁棒性。5.建立了使用人脸-虹膜-指纹三种生物特征进行融合识别的框架。我们首先利用结合PSO和BPSO优化算法分别对人脸训练集、虹膜训练集和指纹训练集的样本进行Gabor核参数寻优,获得提取人脸特征的Gabor滤波器组、提取虹膜特征的Gabor滤波器组以及提取纹指纹特征的Gabor滤波器,它们根据不同的生物模态可自适应确定提取的纹理信息频带。对这些Gabor滤波器提取到的局部化幅值特征和局部化相位特征,用基于SVR融合方法进行特征融合分类,即可实现多种生物特征进行融合识别。实验结果表明,本文方法可以综合人脸-虹膜-指纹三种生物特征的Gabor局部特征进行个体的分类判别,因此识别效果优于单一生物特征的方法和其他融合策略6.提出了使用深度学习中的DBN,通过大样本训练Gabor特征向量的生成模式。在进行虹膜特征提取时,首先使用PSO和BPSO优化算法进行优化后的Gabor滤波器对虹膜图像进行滤波,并对Gabor滤波结果进行相位编码后,将其输入多层全连接的DBN网络中。在经过一定次数的迭代后,将获得相对现有预定义的统计方法获得的Gabor特征向量更复杂模式的自学习特征。在这些复杂模式的虹膜特征中,保留了更丰富的特征表达形式和信息量,因此能更好地获得稳定的虹膜识别效果。对比实验证明,本文提出的最优Gabor滤波器能够产生更为独特的Gabor系数并且其学习特征更具鉴别能力和鲁棒性。此外,本文还探讨了深度学习架构的深度与规模。7.提出通过仿生识别方法进行虹膜识别,这种方法从同类样本的连续性出发,在高维特征空间建立每一类虹膜图像的连续、封闭的最佳几何超体,用来表示每一类虹膜的特征分布。在对待识别图像进行分类时,只需观察其在特征空间中被哪一类的超体所覆盖。这种方法的出发点是对虹膜特征的认知,而不是对虹膜特征的划分,更贴近人类识别新事物的方式,因此被称为仿生识别方法。在自主研发的含视频序列JLUBR-IRIS虹膜数据库上进行了验证实验,展示了本算法在基于虹膜视频序列的虹膜识别系统上的有效性和可靠性。此外,在公共CASIA-I和CASIA-V4数据集的比较实验结果表明,如果提供足够多的样本参与几何超体的训练,本方法还能改善基于静态图像的虹膜识别系统的性能。更多还原

【Abstract】 Gabor filters are generally regarded as the most bionic filters corresponding to the visual perception of humankind. Their filtered coefficients thus are widely utilized to represent the texture information of irises.In this paper, we focus on the domain of iris texture representations and matching, and obtain the following achievements:1. In our iris system, a Particle Swarm Optimization and Boolean Particle Swarm Optimization based algorithm is proposed to train better suitable Gabor filters for each involved test dataset without predefinition or manual modulation. Our system has the advantages of adaptively tuning Gabor parameters, embedded richer informative texture in features. Some comparative experiments on JLUBR-IRIS, CASIA-I and CASIA-V4-Interval iris datasets are conducted, whose results show that our works can generate more excellent local Gabor features by optimized Gabor filters for each dataset.2. This paper has introduced an iris recognition system via multiple local Gabor feature extraction. This system uses two types of Gabor features generated by dividing Gabor responses magnitude and phase to represent iris. The local Gabor response magnitude is the model of orientation for the selective neuron in the primary visual cortex, while the local Gabor phase can capture the information from the wavelet’s zero-crossing3. This paper provides multiple feature representations and their fusion scheme based on Support Vector Regression for iris recognition. All matching scores from multiple local Gabor features are sent into a trained SVR model, and are mapped to a single scalar score to make the final decision. The output of SVR obtained a lower value demonstrate that the test iris and the involved enrollment are in a same pattern class. In light of this principle, a reasonable threshold should be chosen to make the classification decision. Our score fusion by SVR model is superior to other single feature methods in terms of both DI and ROC curves.In this paper, we also compare proposed method with other algorithms and prove its validity and superiority.4. In this paper, we try to improve the characteristics of bionic Gabor representations of each iris via combining the local Gabor features and the key-point descriptors of scale invariant feature transformation (SIFT). A SIFT key point selection strategy is provided to remove the noises and probable misaligned key points. For the combination of these iris features, we propose a support vector regression based fusion rule, which may fuse their matching scores to a scalar score to make classification decision. The experiments on three public and self-developed iris datasets validate the discriminative ability of our multiple bionic iris features, and also demonstrate that the fusion system outperform some state-of-the-art methods.5. Multi-modal biometric system has been considered as a promising technique to overcome the defects of uni-modal biometric systems. In this paper, we have introduced a fusion scheme to gain a better representation and fusion way for face-iris-fingerprint multi-modal biometric system. In our cases, we use Particle Swarm Optimization to train a set of adaptive Gabor filters in order to achieve proper Gabor basic functions for each modality. For closer analysis of texture information, two different local Gabor features for each modality are produced by the corresponding Gabor coefficients. Next all matching scores of the two Gabor features for each modality are projected to a single scalar score via a trained Supported Vector Regression model for final decision. A large-scale combined dataset is formed to validate proposed scheme using FERET-fafb and CASIA-V3-Interval together with FVC2004-DB2a datasets. The experimental results demonstrate that as well as achieving further powerful local Gabor features of multi-modalities and obtaining better recognition performance by their fusion strategy, our architecture also outperforms some state-of-the-art individual methods and other fusion approaches for Face-Iris-Fingerprint multi-modal biometric systems.6. In this paper, an adaptive Gabor filter selection strategy and deep feature learning scheme are presented. We first employ Particle Swarm Optimization rule and its binary version to determine a set of data-driven Gabor kernels for detecting the most informative filtering bands of iris samples involved. Moreover, the further adaptive learning features are also generated by a trained deep belief network to capture complex pattern from the optimal Gabor filtered coefficients. A succession of comparative experiments validate that our optimal Gabor filters produce more distinctive Gabor coefficients and our learned features be more robust and stable on the public CASIA-V4-Interval dataset, CASIA-V4-Lamp dataset and our self-developed larger scale JLUBR-IRIS dataset. Furthermore, the depth and scales of the deep learning architecture are also discussed in this paper.7. In video sequence-based iris recognition system, the challenge of multimodal fusion to make much of relationship and correlation among frames to help recognition still remains to be solved. A brand new template level multimodal fusion algorithm inspired by human being cognition manner is proposed. In that a connected, successive and non-isolated geometrical manifold, named Hyper Sausage Chain due to its sausage shape, is trained at feature space for representing an iris class as a template. The manifold can be utilized to recognize an input iris by observing it in the coverage of the manifold or not. This process is closer to the function of human being to cognize a new object, as its basic principle takes’matter cognition’ instead of ’matter classification’. The experiments on self-developed JLUBR-IRIS dataset including several video sequences per person, CASIA-I and CASIA-V4-Interval demonstrate the effectiveness and usability of our proposed algorithm. Furthermore, our method also can achieve improved performance in image-based iris recognition system provided enough samples involved in training.更多还原