基于多级显著性纹理特征的虹膜图像半监督聚类与分类研究

The personal identification based on the iris recognition receives widespread attention because of the high uniqueness, stability, and non-invasive characters of the iris. The traditional iris recognition technology solved the problem of the matching. However, as the basis of spread of iris recognition, the research on iris image classification is still in its infancy, and different kinds of iris image classification problems (i.e. iris indexing, liveness detection, race classification, forensic information analysis, etc) are isolated from each other which presents limitations. Regarding the unified framework of iris classification as the breakthrough point, this project studies the modeling of iris classification problems, robust texture feature extraction, and iris feature clustering and classification. The research content includes: ①Study the parametric modeling for iris classification task and priori knowledge, data quality assessment and filter, and potential relationship mining; ②Study multi-level saliency of iris image (including database statistical level, object component level, texton level), and propose multi-level saliency based texture feature extraction algorithm; ③Study the semi-supervised sparse subspace clustering algorithm and propose effective similarity measure for iris image clustering, and design a clustering and classification collaboration strategy. Based on the theoretical research, we will develop an iris classification and recognition prototype system. Research results provide theoretical and technical support for the organic combination of the iris classification and the iris recognition system, and can improve the efficacy and safety of the system. It is of great significance for the promotion of large scale person identification base on the iris modal.

由于虹膜的高唯一性、稳定性及非侵犯性特点，基于虹膜的身份认证技术备受关注。传统的虹膜识别解决了比对的问题。但是，作为虹膜识别大规模应用推广的基础，虹膜分类的研究处于起步阶段，不同问题（虹膜检索、活体检测、人种分类等）的孤立研究存在局限性。本项目以虹膜分类的统一框架为切入点，研究虹膜分类问题的建模、鲁棒纹理特征提取、虹膜特征聚类及分类方法。研究内容包括：①研究虹膜分类任务及先验知识的参数化建模，评价筛选数据并研究潜在关系挖掘；②研究多级别显著性（数据库统计级、对象部件级和基元级）对虹膜分类的影响，提出基于多级显著性的虹膜纹理特征提取算法；③研究半监督稀疏子空间聚类算法，提出有效的虹膜相似性测度，设计相互协作的聚类与分类方案。在理论研究基础上，开发虹膜分类与识别原型系统。研究成果为虹膜分类与虹膜识别系统的有机结合提供理论及技术支持，提高系统的效率及安全性，对促进虹膜模态大规模推广具有重要意义。

本项目面向虹膜识别系统应用中对虹膜分类的迫切需求，深入挖掘和研究虹膜分类问题的建模、鲁棒纹理特征提取、虹膜特征聚类及分类方法。主要研究内容包括：（1）研究虹膜分类任务及先验知识的参数化建模，评价筛选数据并研究潜在关系挖掘，分析具有事实依据的虹膜特征与分类关系；（2）研究多级别显著性（数据库统计级、对象部件级和基元级）对虹膜分类的影响，提出基于多级显著性的虹膜纹理特征提取算法；（3）研究半监督稀疏子空间聚类算法，提出有效的虹膜相似性测度，设计相互协作的聚类与分类方案；（4）采用Hadoop分布式文件系统实现分布式并发虹膜分类与识别，验证可行性和便于分类和识别研究过的验证实验。研究成果为虹膜分类与虹膜识别系统的有机结合提供理论及技术支持，提高系统的效率及安全性，对促进虹膜模态大规模推广具有重要意义。