Efficient BSIF-based near-infrared iris recognition

Abstract

Binarized statistical image features (BSIF) represents a general purpose texture descriptor originally designed for texture description and classification, such as local binary patterns (LBP) or local phase quantisation (LPQ). Recently, BSIF has extensively been applied for the purpose of biometric recognition, for instance based on face or palmprint images. While recognition accuracy reported for different biometric characteristics indicates its applicability to iris recognition, up till now BSIF has primarily been employed for iris spoofing detection in particular, fake contact lens detection. In this work we present an adaptation of BSIF for near-infrared iris recognition. In accordance with generic iris recognition schemes, a specific alignment procedure is introduced in order to achieve robustness against head tilts. Further, we propose a binarization method for BSIF-based feature histograms, to obtain a compact feature representation, which allows for a rapid comparison. On the CASIAv4-Interval iris database the proposed system achieves competitive biometric performance obtaining EERs below 0.6%, compared to traditional schemes based on Log-Gabor and quadratic spline wavelets revealing EERs of approximately 0.4%. Moreover, we show that BSIF-based feature vectors complement those extracted by traditional systems, yielding a significant performance gain in a multi-algorithm fusion scenario resulting in an EER below 0.2%, which further underlines the usefulness of the presented approach.