CASIA OpenIR  > 毕业生  > 博士学位论文
异质虹膜识别方法研究
刘年丰
学位类型工学博士
导师谭铁牛
2017-05
学位授予单位中国科学院研究生院
学位授予地点北京
关键词虹膜识别 异质虹膜识别 深度学习
其他摘要

虹膜识别是面向国家战略的重大需求的高精尖技术,现在已经较为广泛的应用在银行、公安和社保等关乎国计民生的行业中。对用户状态和成像设备双重约束的传统的配合式、近距离的虹膜识别系统已经较为成熟, 但是这约束着虹膜识别系统的潜力,使其无法适应与现实世界复杂多变的场景。虹膜图像的来源呈现的多元化趋势,使得如何处理来自不同源的异质虹膜图像识别变得尤为重要。本文围绕异质虹膜识别,使用基于数据驱动的方法学习虹膜的表观特征、数据源之间的空间关系和映射关系、数据中的群体属性和个体联系。并开展了以下的研究:

 

提出了基于层级卷积神经网络模型和基于全卷积神经网络的分割模型来解决非可控场景下采集的低质量虹膜图像的分割问题。这两个模型不依赖于任何先验知识,直接从数据中学习虹膜像素的本征信息。它们属于端到端的模型,优于传统的模型驱动的方法而不依赖任何的前处理和后处理操作;同时,模型的分割精度大大的超过了之前的传统方法。基于全卷积网络的深度分割模型融合了浅层的细节特征和深层的全局特征,因而使得分割精度大为提升。基本解决了非可控场景下虹膜图像预处理问题,提高了虹膜识别系统的用户体验,为异质虹膜识别系统的预处理提供了保障。

 

 

 提出了一种基于图像层的成对深度卷积神经网络的模型(DeepIris),通过学习异质虹膜图像之间的空间关系和表观差异来解决异质虹膜识别的难题。DeepIris通过独特的成对输入设计,使得网络能够自动学习异质源之间的相似性。DeepIris是一个端到端的模型,意味着我们不需要再手工设计特征,即使遇到新的异质源,也可以用DeepIris自动学习出来新老异质源之间的关系。

 

提出了一种新颖的基于编码层的方法来处理多源异质虹膜识别。我们使用马尔科夫模型对异质虹膜图像的二值化编码之间的非线性关系进行建模。通过利用马尔科夫模型学习到的新的编码以及权重,我们的模型可以有效的对抗跨设备,跨采集距离等异质的问题。该编码层方法在跨分辨率,跨设备,跨清晰度等异质虹膜问题上均要好于传统的像素层,特征层和分数层的方法。

 

我们对虹膜图像数据进行系统性的分析,研究基于虹膜图像的群体属性识别和分类问题,包括活体检测、民族、性别等,进一步挖掘群体生物特征群体相似共性和个体关联性,为异质虹膜识别提供新的理论框架和解决方案。我们提出了基于卷积神经网络(CNN)的两个方法来解决虹膜图像分类的问题。其中多子块描述的方法,成功的解决了虹膜图像的异常长宽比对神经网络方法的局限。基于多任务卷积神经网络的方法,通过对虹膜多个属性的相关性来互相提高各自分类任务的准确率。为了研究中国人的虹膜属性问题,我们也专门采集了一个国际上最大的虹膜属性分析数据库。

;

Iris recognition is advanced technology which is important for national strategic agenda. It has been widely applied on banking systems, security departments etc. The study of heterogeneous iris recognition is essential to user-friendly iris recognition systems. This paper try to learn the appearance features, relational relationships, mapping relationships and group attributes from iris data. Our main contribution are summarized as follows:

 

We present a fully convolutional network(FCN) based methods, namely DeepIrisSeg, for both the noisy iris images acquired in visible wavelength and long-distance iris images acquired in near infra-red lights. DeepIrisSeg automatically locate the iris pixels without hand-crafted features or operators, which is an end-to-end model. Rather than using existing model based methods, our model takes the full use of the semantic information from iris images.

 

This paper proposes a deep learning based framework for heterogeneous iris verification, namely DeepIris, which learns relational features to measure the similarity between pairs of iris images based on convolutional neural networks.

DeepIris is a novel solution to iris recognition in two main aspects. 1) DeepIris learns a pairwise filter bank to establish the relationship between heterogeneous iris images, where pairs of filters are learned from two heterogenous sources. 2) Different from two separate steps in terms of handcrafted feature extraction and feature matching in conventional solutions, DeepIris directly learns a nonlinear mapping function between pairs of iris images and their identity supervision with a pairwise filter bank (PFB) from different sources. Thus, the learned pairwise filters can adapt to new sources when given new training data.

 

We proposes a code-level approach in heterogeneous iris recognition. The non-linear relationship between binary feature codes of heterogeneous iris images is modeled by an adapted Markov network. This model transforms the number of iris templates in the probe into a homogenous iris template corresponding to the gallery sample. In addition, a weight map on the reliability of binary codes in the iris template can be derived from the model. The learnt iris template and weight map are jointly used in building a robust iris matcher against the variations of imaging sensors, capturing distance and subject conditions. Extensive experimental results demonstrate the code-level approach can achieve the highest accuracy in compared to the existing pixel-level, feature-level and score-level solutions.

 

We systematically analysis the attributes of the iris data including liveness detection, ethnicity classification and gender prediction. The study of group attributes guide us to building a theory framework and solution to iris area. Two methods are proposed to solve iris image classification. The sub block descriptors successfully break the limitation of unusual length-width ration of normalized iris images when adapting convolutional networks. The multi-task model could predict multiple attributes in one network. We show that the attribute is improving each other when they are learned together. A world biggest database is built to study iris attribute of Chinese.

文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14806
专题毕业生_博士学位论文
作者单位中科院自动化研究所
推荐引用方式
GB/T 7714
刘年丰. 异质虹膜识别方法研究[D]. 北京. 中国科学院研究生院,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
Thesis.pdf(10787KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[刘年丰]的文章
百度学术
百度学术中相似的文章
[刘年丰]的文章
必应学术
必应学术中相似的文章
[刘年丰]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。