Iris Recognition via Non-Local Region Association and Probability Modeling

Pei-hua LI,Guo-long WU,Hong-wei MA

DOI: https://doi.org/10.3969/j.issn.0372-2112.2014.12.007

2014-01-01

Abstract:The traditional iris recognition methods focused on extracting and matching local region characteristics,yet failing to consider the correlation of long-distance regions(non-local regions).The ordinal measures-based methods extracted the average gray-level values of non-local regions by Gaussan filtering and then ranked them,which unfortunately were not applicable to situa-tions where regions are described by probability distributions .This paper presents a novel recognition method for tackling these shortcomings .The proposed method represents the texture characteristics of local regions using spatio-phase joint probability distribu-tions,and further explores the associating relations between them .We concatenate the features of regions at varying positions to ob-tain the non-local associating descriptors for modeling their relations,among which we primarily study the recognition performance by associating two or three regions .During iris matching process,we consider the effectiveness of non-local region associating de-scriptors to exclude the effects of occlusion,highlights,noise and etc,and allow small translation of the associating descriptors to model local deformation of iris texture;finally we adopt a robust diffusion distance between histograms for descriptors comparison . We conduct experiments for iris verification and iris identification on three public databases,and experiments demonstrate that the proposed method is superior to the state-of-the-arts .