ON A LOCAL ORDINAL BINARY EXTENSION TO GABOR WAVELET-BASED ENCODING FOR IMPROVED IRIS RECOGNITION Jinyu Zuo, Natalia A. Schmid Lane Department of Computer Science and Electrical Engineering, West Virginia University, Morgantown, WV 26506, USA jzuo@mix.wvu.edu, Natalia.Schmid@mail.wvu.edu ABSTRACT Daugman’s iris recognition algorithm introduced in early 90s and later undergoing continuous refinements remains poten- tially the most efficient and scalable in iris field. The encoding part of the algorithm relies on application of Gabor wavelets that in terms of their imaging capabilities mimic capabilities of human eye receptor field. In this work, we design and test an algorithm that can be used both individually and as a natural extension scheme to Gabor wavelet-based algorithm. It is based on the local or- dinal information extracted from original unfiltered images. This scheme holds a number of promises: (1) it is robust with respect to a number of nonidealities in iris images and (2) because of the binary nature of the local ordinal information this scheme can be flawlessly integrated into the traditional filter-based recognition systems. The proposed scheme was extensively tested individually and when combined with Ga- bor wavelet-based approach. 1. INTRODUCTION Among existing iris recognition techniques, Daugman’s 2D Gabor wavelet-based encoding algorithm [1, 2, 3] remains the most popular and efficient technique for a frontal view high quality iris imagery. However, since introduction of new re- search directions such as iris at a distance and non-ideal iris the capabilities of Daugman’s algorithms applied to non-ideal iris images were questioned. While most researchers focus on designing novel preprocessing, iris unwrapping, and encoding procedures for non-ideal iris (including Daugman himself), in this work we show that by combining Daugman’s approach with ordinal binary coding at the matching score level leads to considerable recognition gain, efficiency and robustness of a traditional Gabor wavelet-based algorithm. Local Binary Pattern (LBP) technique [4, 5], a subcase of ordinal binary encoding, was previously introduced as an iris encoding technique by Sun et al. [6]. While Sun et al. propose to use LBP as an individual encoding technique, our This work was supported by a grant from NSF IUCRC Center for Iden- tification Technology Research. work pursues different goals. The primary goal of our work is to explore the possibility of use local information contained in a neighborhood of a pixel in images to improve perfor- mance of Gabor wavelet-based techniques. This information can be used to refine computation of Hamming Distances be- tween images and make the combined algorithm more noise resistant compared to the traditional Gabor wavelet-based ap- proaches. The rest of this paper is organized as follows. Section 2 describes ordinal binary encoding and our fusion methodol- ogy. Section 3 evaluates performance of the proposed schemes using both synthetic and real data. Finally, Section 4 presents summary and conclusions. 2. ORDINAL BINARY CODING AND COMBINING METHODOLOGY Consider two iris templates I and I ′ encoded to provide two binary iris codes C I Gabor and C I ′ Gabor . To evaluate match or non-match using a traditional Gabor wavelet-based approach, the codes have to be initially aligned, and the total Hamming Distance (HD) normalized by the number of unoccluded pix- els common to two iris codes is calculated. The alignment part as well as matching part can be easily refined. Since iris code contains information about both real and imaginary parts, for a certain pixel pair, for example, the pixel p 0 from templates I and the pixel p ′ 0 from templates I ′ , HD can only take three possible values: 0, 1 and 2, since we have 2 bits to describe a value of each pixel. The two bits for pixel p 0 are ℜ p 0 and ℑ p 0 . The two bits for pixel p ′ 0 are ℜ p ′ 0 and ℑ p ′ 0 . To provide an example we list all possible combinations and values that HD can take in Table 1. From here the normalized HD for each pixel pair can take only three values: 0, 0.5 and 1. It can be interpreted as following: 0: Two pixels with their surroundings are similar. 0.5: The relation between two pixels is uncertain. 1: Two pixels with their surroundings are different.