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 situations 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 distributions

and further explores the associating relations between them.We concatenate the features of regions at varying positions to obtain 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 descriptors 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.