J.

Daugman

New methods in iris recognition

IEEE Trans. Systems. Man. Cybernetics B, 37 (5) (2007), pp. 11671175

Abstract

This paper presents the following four advances in iris recognition: 1) more
disciplined methods for detecting and faithfully modeling the iris inner and outer
boundaries with active contours, leading to more flexible embedded coordinate
systems; 2) Fourier-based methods for solving problems in iris trigonometry and
projective geometry, allowing off-axis gaze to be handled by detecting it and
"rotating"the eye into orthographic perspective; 3) statistical inference methods for
detecting and excluding eyelashes; and 4) exploration of score normalizations,
depending on the amount of iris data that is available in images and the required scale
of database search. Statistical results are presented based on 200 billion iris cross-
comparisons that were generated from 632500 irises in the United Arab Emirates
database to analyze the normalization issues raised in different regions of receiver
operating characteristic curves

Abstract
Ocular biometrics encompasses the imaging and use of characteristic features extracted from
the eyes for personal recognition. Ocular biometric modalities in visible light have mainly
focused on iris, blood vessel structures over the white of the eye (mostly due to conjunctival
and episcleral layers), and periocular region around eye. Most of the existing studies on iris
recognition use the near infrared spectrum. However, conjunctival vasculature and periocular
regions are imaged in the visible spectrum. Iris recognition in the visible spectrum is possible
for light color irides or by utilizing special illumination. Ocular recognition in the visible
spectrum is an important research area due to factors such as recognition at a distance,
suitability for recognition with regular RGB cameras, and adaptability to mobile devices.
Further these ocular modalities can be obtained from a single RGB eye image, and then fused
together for enhanced performance of the system. Despite these advantages, the state-of-the-
art related to ocular biometrics in visible spectrum is not well known. This paper surveys this
topic in terms of computational image enhancement, feature extraction, classification schemes
and designed hardware-based acquisition set-ups. Future research directions are also
enumerated to identify the path forward.

Keywords

Biometrics;

Ocular biometrics;

Mobile biometrics;

Iris;
 Conjunctival vasculature;

Periocular biometrics;

Visible spectrum

1. Introduction
Biometrics is the science of establishing identity of individuals using physical (such as face,
fingerprint and iris) or behavioral traits (such as voice and gait) [58] ; [123]. Among others,
significant advancement has been observed in ocular biometrics over the past few years.
Ocular biometrics is a sub-field of biometrics focused on regions in the eye and those around
it. Iris has traditionally been considered as the most popular and widely accepted ocular
biometrics due to its high accuracy and stability over individual's lifetime [36]; [75]; [78] ;
[93]. Virtually all existing commercial iris recognition systems operate in near-infrared (NIR)
illumination, and many use constrain acquisition protocols in order to minimize the impact of
intra-class variations due to pose, distance, illumination changes, and occlusion from
eyeglasses, eyelashes and specular reflections [79].

Academic research groups have reported some advances in iris recognition under visible
spectrum (VIS) [97] ; [103]. However the richness of iris texture is not always well captured
in a VIS image due to melanin content in darker iris images. In fact, studies suggest that NIR
iris images significantly outperform VIS iris images from a matching accuracy
perspective [20]. Nevertheless, iris recognition in the visible spectrum may facilitate
recognition when NIR iris captures are lacking or inadequate e.g., due to distance. Further,
VIS iris captures can potentially assist in classification and indexing based on iris macro
features (e.g., moles, freckles and nevi), and eye color [93] ; [104].

Apart from iris, human eyes carry other distinguishable patterns such as conjunctival
vasculature [41][28][35], periocular [93] and retinal biometrics [47] which have been used
for personal recognition. Fig. 1 shows an RGB eye image with its iris, conjunctival
vasculature and periocular region. The blood vessel from the conjunctiva and episclera seen
atop of the avascular white of the eye (sclera), and henceforth called conjunctival vasculature
for brevity, is distinct to each individual, and can be non-intrusively acquired in the visible
wavelengths [28] ; [41].

Fig. 1.

An example eye image labeled with iris, conjunctival vasculature, and periocular
region.
 Figure options

The feasibility of conjunctival vasculature as a biometric has been well established in the
literature [16]; [41]; [93] ; [95]. The feature size resolved by any optical system (spatial
resolution) is proportional to the wavelength used. Thus, green and blue spectrum (used for
conjunctival vasculature) have better resolving power compared to the NIR used for iris
recognition [41]. The formula (Rayleigh's) for spatial resolution is as follows:

equation(1)

l=1.22fD
Turn MathJax on

where f is the focal distance, is the light wavelength, and D is the aperture diameter. Thus,
given the NIR, green, and blue wavelengths of 850 nm, 550 nm, and 450 nm, respectively,
optically, conjunctival vasculatures' resolving power is 55% to 89% higher than iris captured
in NIR at the same distance [28]; [35] ; [41]. In other words, conjunctival vasculature visible
light scans can achieve the same spatial resolution at farther distances compared to NIR-based
iris recognition system, other imaging aberrations not withstanding.

Periocular biometric refers to area around eyes such as eyelids, eye-folds and even
eyebrow [7]; [16] ; [93]. In some publications, the term has been loosely used to refer to the
entire ocular region (or the whole eye image) [16]; [39]; [89] ; [93]. Studies have also
combined periocular region with the degraded iris data to enhance the performance of the
overall system in visible light [91] ; [101]. Retinal biometrics cannot be captured using
regular RGB cameras and specialized optical devices designed specifically to capture the back
of the eye is needed for this modality [41]. Further, retinal biometrics may not have high user
acceptability because they require high degree of user cooperation and are considered to be
invasive. Thus, retinal biometrics is not further discussed in the paper.

All these ocular traits i.e., iris, conjunctival vasculature, and periocular region which (except
for dark irises) can be potentially captured from a single RGB eye image are referred to as
ocular biometrics [26] ; [93]). Moreover, these ocular traits can be fused together for
enhanced performance of the overall system [90]; [112] ; [114] which is referred to as intra-
ocular fusion.

Despite the aforementioned advantages associated with ocular biometrics in the visible
spectrum, the low quality of images captured in unconstrained conditions (i.e., uncontrolled
environment) remains to be a challenge. This is due to the factors such as motion blur, eyelid
occlusion, eyelashes, specular reflections, gaze deviation, pose variations and challenging
illuminations. Further, RGB eye images may demonstrate low contrast between the iris and
pupil boundary, especially for a dark pigmented irides [110]. This is because photic
accumulation in RGB is so low that texture may be lost in sensor noise. This problem is
exacerbated for darker irises given the IR cut filters and lower signal-to-noise ratio (SNR) of
RGB, especially R compared to G in Bayer patterns. Consequently, lack of texture, incorrect
localization, and error in segmentation leads to lower RGB iris recognition accuracy [110].
The impact of some of these factors may be mitigated by adopting effective computational
image enhancement, robust feature extraction and classification schemes. The image
enhancement schemes can (a) enhance the signal-to-noise ratio i.e., bringing out the details,
such as rendering pupil or conjunctival vasculature pattern [41] more visible, and (b) mitigate
some of the artifacts, such as specular reflection and motion-blur. The enhanced eye images
help in accurate segmentation, feature extraction, classification, and hence, higher recognition
accuracy. An alternative is to design better hardware set-up to acquire relatively high quality
RGB ocular data, especially when it comes to illumination [51] ; [145]. Therefore, the
problem associated with RGB ocular biometrics in an unconstrained environment and setup
can be divided into two major sub-problems: (1) hardware design for obtaining relatively
good quality ocular data, and (2) recognizing individuals using degraded ocular data by
applying robust image enhancement, feature extraction and classification schemes.

This paper reviews RGB ocular biometrics literature in terms of computational image
enhancement, segmentation and feature extraction schemes for iris, conjunctival vasculature
and periocular region captured at VIS. Hardware-based acquisition set-ups designed to
acquire high quality ocular data are discussed as well. Further, future avenues to identify the
path forward in this field are also enumerated.

The specific contributions of this paper are enlisted as follows:

1.

A survey of ocular biometrics at VIS in terms of computational image enhancement,

feature extraction and classification. Related hardware-based set-ups are discussed as
well.

2.

Enumeration of publicly available ocular biometric databases captured at VIS.

Tabulation and analysis of the results as reported by the researchers on these datasets.

3.

Discussion of the future research directions for ocular biometrics at VIS.

Nigam et al. [89] reviewed the advancements made in iris and periocular biometric
modalities. The advancements and developments were measured in terms of the technical
contributions to enhance the recognition performance, and collection of databases that help in
addressing important emerging challenges. However, the survey was presented at NIR
spectrum, and, VIS has not been discussed in detail. Further, computational image
enhancement schemes adopted for ocular biometrics has not been discussed. Conjunctival
vasculature biometrics has not been included. In contrary, this paper exclusively focuses on
surveying ocular biometrics in the visible spectrum. In addition to iris and periocular region,
survey on conjunctival vasculature has been included. Further, computational image
enhancements for noisy ocular images captured at VIS (for improving recognition accuracy)
have been discussed.
The rest of this paper is organized as follows: Sections 2 and 3 survey iris and conjunctival
vasculature in the visible spectrum in terms of image enhancement, features extraction and
classification schemes adopted for ocular biometrics. Section 4 discusses periocular
biometrics. Hardware-based schemes for improved acquisition set-ups to obtain better quality
ocular data are discussed in Section 5. In Section 6, publicly available databases on ocular
biometrics are enlisted. In Section 7, results of the existing VIS recognition schemes based on
iris, conjunctival vasculature, periocular and their intra-ocular fusion have been discussed.
Section 8 presents discussion and possible future research directions.

2. Iris biometrics in visible spectrum

Iris recognition is a method of personal authentication based on distinguishable and unique
patterns captured from the texture of individual's iris [38]; [78] ; [104]. Several hundred of
millions of people are authenticated using iris recognition technology for security applications
involving border crossing (such as OBIM program1), passport control (United Arab Emirate's
immigration program2) and national ID program (such as UIDAI3).

An iris recognition system is typically composed of (i) image acquisition, (ii) segmentation
and normalization, and (iii) feature extraction and matching [105] ; [152]. Iris recognition
works by acquiring the eye image, localizing the pupil and excluding the noisy regions. This
step is called iris segmentation, and is considered very important as incorrect segmentation
will negatively impact iris recognition performance. However, effective iris area can vary
significantly across eye images due to photic pupil dilation and contraction, sleepiness or
alcohol consumption [10]. User distance, sensor resolution, and angular magnification adds to
this variability. To address the variation in the iris band size, the segmented iris is unwrapped
and mapped to a polar coordinate system. This step is called iris normalization. One of the
most popular techniques for iris normalization is the Daugman's rubber sheet model [37] that
maps every point in the segmented iris region from cartesian to polar coordinates. Finally,
features are extracted to encode discriminatory information from a pair of normalized iris
images and are matched to authenticate an identity.

Typically, iris recognition systems operate under NIR illumination with wavelengths ranging
between 700 nm to 900 nm. Under NIR illumination, the effect of melanin is negligible.
Therefore, operating at NIR spectrum ensures that the acquired image reveals information
related to the texture rather than pigmentation. The nuances of the iris texture of dark-colored
irides are much better observed in the NIR spectrum [56]. However, under special lighting
arrangements or with light colored irides, iris recognition at VIS is doable.

Iris recognition at VIS offers some advantage over NIR as follows:

Recognition at a Large Distances: VIS could be more amenable to light-colored iris

recognition at a distance or on-the-move applications, assuming ample lighting and
absence of detrimental specularities and glare. This is because for iris recognition at a
large distance, acquisition environment needs to be well illuminated and natural and
adequate NIR illumination is not usually present [150].

 Affordable and Widely Available Cameras: Compared to NIR, VIS sensors are
relatively cheaper and much more prevalent thanks to popularity of mobile phones.

Classification and Indexing: Color of the eye can be obtained from iris images
captured at VIS. Such information together with iris macro-features (e.g., moles,
freckles and nevi) can facilitate iris classification and indexing [137]. Degraded iris
images have been indexed utilizing iris code obtained after feature extraction and
encoding process [100].

However, studies indicate that performance of the iris recognition system significantly
degrade under VIS [20]; [104] ; [142]. In fact, an international competition called Noisy Iris
Challenge Evaluation- Part II (NICE.II) [20] was conducted with an aim to evaluate the
performance of the iris recognition under VIS rather than NIR. The results of the NICE.II
competition [20] suggest degradation in the performance of the iris recognition at VIS over
NIR.

The performance improvement in iris recognition at VIS can be leveraged by (i) robust image
enhancement, feature extraction and classification schemes: in order to mitigate the impact of
intra-class variations, such as defocus, motion blur, off-angle imaging, illumination variations,
and specular reflection, which adds noise to the acquired iris image [81]. Fig. 2 shows
example low quality eye images acquired in the visible spectrum, (ii) information fusion: the
top performing algorithms in NICE.II competition [20] showed that fusion of the features
from the iris region with those extracted from other ocular regions (conjunctival vasculature
and periocular) significantly improves the performance. Another alternative for improved
performance is the part-based fusion where the features or matching scores from multiple
regions of the ocular trait are fused together to deal with misalignment or inaccurate
segmentation, and (iv) improved hardware set-up: especially lighting to obtain better quality
ocular data.
 Fig. 2.

Example low quality eye images captured in visible spectrum. The low quality is due
to factors such as motion blur, occlusion and low contrast between pupil, iris and
sclera (white of the eye).

Figure options
2.1. Methods for iris image enhancement

In this section, we will discuss computational image enhancement schemes adopted for iris
biometrics in the visible spectrum. In most of the existing studies, these image enhancement
schemes have been applied for contrast enhancement and illumination normalization,
deblurring, and specular reflection removal [110] in RGB iris images as discussed below.
However, a strong pre-processing scheme may also induce non-existing features to the image
given pre-processing bias, and increase intra-subject variability. For instance, applying phase
congruency on a noisy scleral image will yield non-existing features for conjunctival
vasculature patterns.

Contrast Enhancement and Illumination Normalization: Santos and Hoyle [127]

applied adaptive histogram equalization scheme on an eye image captured under low
lighting to facilitate detection of pupil boundaries in the search area. In adaptive
histogram equalization, histograms are computed within non-overlapping tiles of an
image to redistribute pixel intensities. Raffei et al. [110] proposed ISO-contrast limited
 adaptive histogram equalization to enhance the contrast of the eye images captured
under low lighting or with low contrast ratio between the iris and pupil boundary.

Jobson et al. [61] and Shin et al. [130] have applied the Retinex algorithm to reduce
illumination variation in eye images. This algorithm is based on decomposing a given
image into two images: the reflectance image and the illumination image. The
reflection image represents the unique characteristics of the image after eliminating
the illumination part. Shukri et al. [131] used homomorphic filtering followed by
multi-scale Retinex to enhance the details in an eye image. Homomorphic filtering
involves nonlinear mapping of the image to a different domain and applying linear
filters to the image in transformed domain. The filtered image in transformed domain
is then mapped back to the original domain.

Deblurring: Optical and motion blurring is quite common especially in mobile iris
recognition and iris-on-the-move applications [80] due to user and camera
movements. De-focus or optical blurring may also happen due to non-optimal focal
length, depth-of-focus, lens aberrations, or auto-focus systems or metering issues.
In [64], the implemented motion and optical de-blurring procedure consisted of three
stages: motion blur assessment, optical blur assessment, and image restoration. To
facilitate this, direction and magnitude of the motion and focus score, the amount of
high frequency components in the image, were estimated. Finally, based on the
magnitude of the motion and the focus score, it was decided if the image restoration
will be performed or not using modified Constrained Least Square filter.

Specular Reflection: The presence of specular reflection may be detected by analyzing

pixel intensities through thresholding. However, it is very difficult to define an
optimum threshold value. To overcome this problem, Lee et al. [69] proposed
threshold method of line intensity profile (LIP) for reflection detection. In this method,
a pixel was labeled as a reflection if its red intensity value was higher than its blue and
green intensities. Two sets of pixel intensity vectors from reflection and non-reflection
areas as identified by LIP, were used to train a two-class SVM. After reflection
classification using the trained SVM, it was removed by interpolation using nearby
pixel values lying outside the reflection area [108]; [109] ; [125].

Studies in [72] ; [141] proposed schemes to estimate noise from the iris images. Li and
Savvides [72] used Gaussian Mixture Model (GMM) to model the distribution of regions in
iris images for robust occlusion estimation. Tan and Kumar [141] derived a noise model from
the temporal information by exploiting the consistency and the similarity of the intra-class iris
codes. The bits which were consistent over time were assigned higher weights during the
matching process. Iris segmentation methods for mobile applications have been reviewed
in [60]. Recently, histogram of oriented gradients combined with SVM have been used for iris
segmentation [57]. Iris detection and tracking at VIS have been done using Watershed [1],
Recursive Prediction Analysis [52] and Mean-shift algorithm [22].
2.2. Methods for iris image representation and classification

Existing iris image representation methods can be classified as follows:

Global vs Local Texture Descriptors: Global features are extracted from the entire
image as opposed to local features which are extracted from multiple local regions in
the image. Barra et al. [14] used spatial histogram (spatiogram) for iris feature
representation. Spatiogram was suggested to efficiently calculate the differences
between images, irrespective of geometric transformation between them. In [70],
weighted co-occurrence phase histogram was used for encoding texture patterns in iris
image by weighing the co-occurrence of pair of image pixels. Further, the
characteristics of joint probability distribution of both the phase angle and spatial
layout resulted in efficient description of texture patterns. A combination of Gabor
Wavelets with SIFT was used to improve iris recognition in visible spectrum [45]. A
non-linear approach to account for local consistency of iris bits and the quality of the
weight map using Zernike moment which penalizes the fragile bits while
simultaneously rewarding more consistent bits was proposed [139]. Studies in [48];
[97]; [104] ; [119] used 2D Gabor Wavelets for feature extraction from color iris
images. Li et al. [48] fused phase and zero-crossing methods for iris template
generation. Log-gabor filters have been used for iris feature extraction in [6] ; [82].
The approach in [20] ; [130] fused iris codes obtained using Gabor Wavelet from the
red, green and gray-scale versions of the iris images. Santos et al. applied 1-D
Gaussian wavelet transform to each row of pixels in normalized and segmented iris
images to obtain a one dimensional periodic signal [127].

Local texture descriptors such as Local Binary Pattern (LBP) [146] and Binary
Statistical Image Features (BSIF) [117] were used for iris feature extraction. LBP
operates by comparing intensity of each pixel to its neighborhood pixels and encoding
the result as a binary bit string. BSIF computes a binary code for each pixel by linearly
projecting local image patches onto a subspace, whose basis vectors are learnt from
natural images via independent component analysis. A method based on tensor
decomposition of SIFT key point descriptors was proposed by tessellation of the iris
image into sub-images or patches [73]. SIFT descriptors were extracted from each
patch and encoded into binary codes by comparison with the mean value.

Pixel(Patch)-based vs Transformed Domain: Tan and Kumar [140] proposed an

algorithm that exclusively assigns a set of coordinate-pairs to each user in the system.
The geometric key uniquely encodes the way in which iris features are extracted from
the localized iris pixels. The iris encoding scheme constitutes of computationally
efficient operations on the image patches based on the locations defined by the
geometric key. K-means clustering have also been used to train a set of filters which
are the K-means centroids learnt from the patches of natural images. These filters were
considered as the units of learned dictionary and were used for feature extraction from
iris images [116].
 Radu et al. [106] calculated Circular Symmetric Filters, which are the product of a
Gaussian envelope and a sinusoidal function, for iris image representation. The Local
Radon Transform was adopted for feature extraction from iris images [108].
Specifically, a multi-scale version of Local Radon Transform was used to obtain
highly precise iridial information at different scales. Study in [101], extracted iris
codes by convolving the normalized iris data, obtained by converting segmented iris
image from Cartesian to polar domain, with a bank of Multi-Lobe Differential Filters
(MLDF) which are expressed in terms of the number of lobes, location, scale,
orientation and inter-lobe distance.

Shape vs Color Features: Hosseini et al. [56] proposed shape features to encode the
pigment melanin patterns in visible light. Specifically, three features were proposed
i.e., radius vector function, support function, and tangent angle function. Proena et
al. [98] combined color and shape features to enhance the performance of the
recognition system.

Demirel and Anbarjafari [53] concatenated histograms of the iris image in Hue,
Saturation and Intensity (HSI) color space. Similarly, a study in [155] explored iris
color texton in different color spaces. The proposed scheme combined pixel values in
RGB, HSI and LAB color spaces where the image was represented by a histogram of
the learned iris color texton vocabulary.

Graph vs Sparse Representation: Farouk et al., proposed a new method for iris
recognition based on Elastic Graph Matching (EGM). In this method, the segmented
irises were represented as labeled graphs. Nodes were labeled with jets and edges were
labeled with distance vectors [46].

Raja et al. [121] used Deep Sparse Filtering, based on unsupervised algorithm to learn
the number of specified features which does not explicitly attempt to model the
distribution of data, for iris image representation. Further, in [68] the combination of
hyper complex and sparse representation was used on segmented and normalized iris
images. The orientation of local iris texture elements was extracted using a binarized
dictionary of oriented atoms.

Combination of texton representation, color analysis, ordinal measures, color histogram, and
semantic information was proposed for iris recognition in visible spectrum [20] ; [142].
These multiple cues were fused at match score level to obtain a final dissimilarity measure.
Santos et al. [127] performed a multi-classifier fusion of iris representation based on (a) both
1D wavelet and 2D wavelet zero-crossing maps, (b) 2D Gabor filters, and (c) LBP and SIFT
features, was proposed. The fusion was performed using logistic regression approach.

Classification: For iris image classification, most the systems relied on hamming distance
metric [20]; [48]; [56]; [97]; [98]; [101]; [104]; [119]; [127]; [130] ; [146], while fewer used
other measures such as cross-correlation [53] ; [155], phase distance function [139] and point
pattern matching [73]. Recent studies used some elaborate techniques based on sparse
representation classification [117] ; [121] and graph matching [46].

3. Conjunctival vasculature
The conjunctiva is a thin, transparent, and moist tissue that covers the outer surface of the
eye [3]; [33]; [35]; [40] ; [158]. The ocular conjunctiva is very thin and clear; thus the surface
vasculature (due to conjunctiva and episclera) are easily visible through it. The visible micro-
circulation of conjunctiva and episclera offers a rich and complex network of veins and fine
micro-circulation [40]. Note that the white of the eye, or sclera, is avascular.

Conjunctival (and episcleral) vasculature confirms to the following basic criteria of acceptable
biometrics [40] ; [41]:

Universality: Every human has conjunctival blood vessel structures on the white of
the eye.

Permanence: Conjunctival vasculature is usually captured when the user is gazing to

the left or right (while fixated, for instance, on a target on the screen of capturing
device) with the gaze direction being coaxial with the capturing camera lens or
looking directly into the camera. Conjunctival vasculature patterns have permanence
within the predetermined gaze angles that remain reasonably consistent between the
enrollment and verification session. However, no gazing solutions are often
challenging because of the smaller scleral ROI in a no gazing user interaction.

Uniqueness: Vasculature patterns are created during embryonic vasculogenesis and

are unique to each individual. Studies conducted in other areas, such as eye fundus,
validate the uniqueness of the conjunctival structures for each individual [40].

Practicality: The practicability of the conjunctival vasculature is high because

commercial-off-the-shelf RGB cameras can be used to efficiently acquire conjunctival
vasculature under daylight, dim or office light settings without any specific
illumination device.

Acceptability: Conjunctival vasculature is highly acceptable. Especially for mobile

use case, capturing conjunctival vasculature is tantamount to taking a selfie of one's
ocular region.

Spoof-proofness: Existing studies claim that fine multi-scale blood vessel structure
make them difficult to spoof [40].
Conjunctival vasculature recognition steps include image processing, segmentation and
enhancement of the vascular region followed by feature extraction and matching. Besides
being a capable stand-alone modality, the fusion of conjunctival vasculature with iris
biometrics [33] may improve the performance as: (a) additional information pertaining to iris
biometrics will be available, (b) in off-angle iris scenario, where the iris is moved to the left or
right, the conjunctival vasculature patterns are better revealed, (c) errors related to failure-to-
enroll for iris, for instance, due to medical conditions such as cataract [113], will be accounted
for, and (d) robustness to spoof attacks will be increased. This is because both the iris and
conjunctival vasculature have to be accurately imitated in the spoofed eye image in order to
fool the biometric system for illegal access and advantages.

In fact, conjunctival vasculature may also offer additional advantages over iris biometrics as
follows:

1.

Conjunctival vasculature biometrics uses green-blue frequencies that have shorter

wavelengths compared to iris at NIR. Thus, conjunctival vasculature patterns can be
optically better resolved at farther distances.

2.

An important issue with iris recognition is contraction and dilation of the pupil [54]
due to several factors such as ambient light, sleepiness, emotional states, or alcohol
consumption [10], leading to performance degradation, a non-issue for conjunctival
vasculature.

3.

Non-prescription cosmetic lenses that obstruct scanning of iris patterns, do not affect
conjunctival vasculature feature extraction and segmentation. Studies have shown that
even clear prescription contact lenses negatively affect iris recognition accuracy [11] ;
[54], yet again, not a problem for conjunctival vasculature.

4.

The available conjunctival vasculature areas per person are four (to the left and right
side of the iris in the left and the right eye), compared to two irises per person. This
provides more flexibility and robustness for conjunctival vasculature based
biometrics [28]; [35] ; [145]. Further, in the presence of common eye occlusions such
as glare and eyelashes, using conjunctival vasculature (by themselves or in fusion with
iris) can reduce the chances of false rejection.

5.

It should be noted that though scientifically unproven, some users may also be
concerned about repeated infrared irradiation of their eyes by iris scanners operating at
invisible NIR [66] ; [147], whereas conjunctival vasculature use the natural ambient
visible light [31].
6.

Further, conjunctival vasculature patterns have more contrast in comparison to iris

patterns because conjunctival vasculature arcades appear almost black in green/blue
wavelengths against their mostly white scleral background [41]; [143] ; [145]. This
means that iris scanner demand higher-end Modulation Transfer Function
characteristics in terms of light/dark percentage in comparison to conjunctival
vasculature. This advantage of conjunctival vasculature also implies that it is less
sensitive to sensor noise, since their higher contrast patterns are less likely to be
degraded due to thermal noise. Therefore, as the image contrast and the amount of
incident light are reduced, conjunctival vasculature scans may yield a better signal-to-
noise ratio (SNR) [30] ; [144].

Next, we will review image enhancement and scleral segmentation schemes adopted for
conjunctival vasculature patterns.

3.1. Methods for conjunctival vasculature enhancement

The aim of image enhancement routine is to further delineate conjunctival vasculature

patterns against their white scleral background. In most of the existing studies [33]; [35] ;
[40], enhancement is usually performed either on the entire acquired eye image or the scleral
region for contrast enhancement, image denoising and specular reflection removal. Then, the
green layer of the enhanced RGB color image is used for the detection and extraction of blood
vessels. This is because green channel exhibits better contrast between the conjunctival
vasculature and the background [41]; [143] ; [144].

Contrast Enhancement: Studies in [26]; [41]; [143] ; [144] have adopted Contrast
Limited Adaptive Histogram Equalization (CLAHE) to enhance conjunctival
vasculature patterns prior to feature extraction and classification.

Study in [156] applied selective line enhancement algorithm for enhancement of the
blood vessel patterns details seen in the green channel. The algorithm locates lines and
curves based on the second derivatives from the Hessian matrix. However, because
second order derivative is sensitive to noise, the eye image is first convolved with a
Gaussian Kernel whose standard deviation is commensurate with the scale of the lines
and the girth of the blood vessel patterns to be detected. The Sobel operator is an edge
enhancement that convolve the input image with integer valued kernel in either
horizontal or vertical direction and creates an output image with enhanced edges and
transitions [74]. Studies in [149] ; [158] used Sobel filter to enhance conjunctival
blood vessel patterns. Studies in [156] ; [157] applied a bank of Gabor filters for
conjunctival pattern enhancement given that vascular patterns have certain
orientations.

Image Denoising: Studies in [26] ; [34] employed double density complex discrete
wavelet transform (DDCDWT) and mayer wavelets, respectively, to denoise
 conjunctival vasculature patterns. In wavelet based image denoising, the input image
is transformed into an orthogonal domain by using a wavelet function (e.g.,
Daubuchies, Symlet, Mayer etc.). Then, soft or hard thresholding is applied to the
wavelet coefficients to suppress smaller amplitudes. Lastly, inverse discrete wavelet
transform is applied to obtain the denoised image [19] ; [44].

Specular Reflection: Crihalmeanu and Ross [26] applied a fast in-painting algorithm
for accurate detection and removal of specular reflection. Such inpainting is to
facilitate segmentation of the scleral region, and cannot recover lost vascular
information due to severe specularities. The algorithm derives information from the
boundary of the to be in-painted region, and smoothly infuses estimated pixel values
inside the specular region using a partial differential equation.

Authors in [28] ; [143] performed segmentation of the Scleral region using k-means
clustering algorithm with Euclidean distance metric. K-means clustering was applied on
image pixels with each pixel representing a 3-dimensional vector of its RGB intensities. Then,
the pixels were classified as belonging to iris, sclera or skin region based on their respective
cluster membership. Derakhshani et al. [41] applied region growing method for segmentation
of conjunctival vasculature. The proposed algorithm is based on classifying each pixel as
blood vessel or background (the white of the eye) in an iterative manner based on the local
intensity values and image gradient magnitude. Authors in [33] used time adaptive active
contour-based method for Scleral segmentation as a precursor for processing conjunctival
vasculature.

3.2. Methods for conjunctival vasculature representation and classification

Existing methods for conjunctival vasculature representation can be classified as follows:

1.

Global vs Local Texture Descriptors: Global descriptors such as Gray-level co-

occurence matrix (GLCM) [143] and Histogram of Oriented Gradient (HOG) [136]
are used for feature extraction of conjunctival vasculature. GLCM operates by
calculating the frequency of pair of pixels with specific values and spatial relationship.
Contrast, correlation, energy and homogeneity features of GLCM matrix were used as
features for conjunctival vasculature [143]. HOG counts occurrences of gradient
orientation in localized portions of an image.

Local descriptors such as Local Binary Pattern (LBP) [32], Binary Statistical Image
Features (BSIF) [90], Local Directional Patterns [90], Scale Invariant Feature
Transform (SIFT) [33] and Speeded-up Robust Features (SURF) [27] were used for
feature extraction. In [33], Dense Scale Invariant Feature Transform (D-SIFT) was
used for conjunctival vasculature feature extraction. D-SIFT patch descriptors of
training images were used to form bag-of-words. In [27], SURF descriptors were used
for conjunctival vasculature feature extraction. Study in [50] combined SURF with
Gabor features for conjunctival vasculature feature extraction.
2.

Pixel-based vs Transformed Space: In pixel space based methods, features are

extracted by direct comparison of the image pixels. Crihalmeanu and Ross [26],
estimated similarity between two sclera images using direct pixel value cross-
correlation between aligned regions of the sclera within the overlapping areas of the
two scleral masks. Zhou et al. [156], used thinning (morphological operation) to
reduce the detected blood vessel patterns to single-pixel wide skeleton. Then, the line
segments, described by three quantities: reference angle at the pupil center, the
segment distance to the pupil center and the dominant angular orientation, were used
as features for conjunctival vasculature. Derakhshani and Ross [41], represented blood
vessel patterns using discrete minutiae points. The minutiae points represented the
locations of vascular branching (anastamosis points).

In transformed space based methods, image is transformed to a different domain

before feature extraction. Derakhshani and Ross [40] used top packet energies of the
wavelet coefficients at different decomposition levels as features. Discrete Cohen-
Daubechies-Feauveau 9/7 was used as a mother wavelet. Hu [84] derived six absolute
orthogonal invariants and one skew orthogonal invariant based upon algebraic
invariants, which are independent of position, size, orientation and parallel projection.
Derakhshani and Ross [41] used moment invariants for coarse-level conjunctival
vasculature matching.

Classification: For conjunctival vasculature classification, most of the existing systems used
distance metrics such as Euclidean distance [41] and hamming distance [50], point-pattern
matching [27] ; [41], and cross-correlation [26]. Few studies adopted learning based
classifiers such as Neural Nets [40] and SVM [31]; [32]; [33] ; [90] for final classification.

The possible pitfalls include occlusion due to droopy eyelids, inconsistent gazing between
enrollment and verification, and an extensive glare due to prescription eye glasses or sun
glasses would hamper scanning and matching of conjunctival vasculature patterns. Further,
conjunctival hemorrhage due to direct trauma to the eye or very rare conditions such as ocular
sarcoma [132] will impact the matching accuracy.

4. Periocular biometrics
The term periocular refers to the region surrounding the eye area such as eyelids, eye-folds
and eyebrows. Various studies have established the potential of periocular region as a
biometric trait [7]; [16] ; [93]. Periocular biometric trait can also be advantageous as a
complimentary modality, when other ocular traits exhibit large intra-class variations. For
instance, periocular region can be used to authenticate an individual when the acquired iris
image exhibit blinking or other occlusions as intra-class variations [7]; [16]; [18]; [91]; [93] ;
[101].

Mostly, global features (such as Gradient Orientation (GO) and Histogram of Oriented
Gradients (HOG) [154]) and local features (such as Local Binary Patterns (LBP) [94], Binary
Statistical Image Features (BSIF) [117], Scale Invariant Feature Transform (SIFT) [127] and
Speeded-up Robust Features (SURF) [154]) have been used to extract features from the
periocular region. Miller et al. [83] demonstrated that it may be beneficial to use the green
channel of an RGB image for periocular recognition.

Periocular region segmentation has been extensively discussed in the existing literature. Park
et al. [94] studied manual and automatic segmentation schemes with and without eyebrows.
Further, the impact of masking/ segmenting the sclera and iris from the periocular region was
also evaluated. Studies in [5] ; [134] investigates what is the best periocular region by
varying the portion of the periocular region, and regions of the periocular region which are
beneficial to recognition. Smereka et al. [135], proposed an unsupervised dynamic
segmentation approach where periocular image is segmented into rectangular regions which
improve the overall discrimination ability of the pair of biometric samples being matched. For
periocular region image enhancement, it is safe to say that many facial image enhancement
methods such as self quotient image (SQI) [151] are applicable to periocular region as well.
Next, we summarize feature extraction and classification schemes for periocular biometrics.

4.1. Methods for periocular representation and classification

Existing feature extraction schemes on periocular representation can be classified as follows:

1.

Global vs Local Texture Descriptor: Global descriptors such as HOG [93] ; [94],
GIST (summarizes the gradient information (scales and orientations) for different parts
of an image) [16], Gabor filters [5]; [6] ; [154] and Color histogram [153] were used
for feature extraction of periocular region.

Local descriptors such as LBP [2]; [12]; [16]; [91]; [93]; [94]; [95]; [127] ; [154],
SIFT [5]; [91]; [93]; [94]; [127] ; [154] and SURF [154] were used for feature
extraction of periocular region.

2.

Spatial vs Transformed Domain: Study in [43] used shape-based features for

periocular feature extraction. The x, y positions of endocanthion and exocanthion were
recorded as reference points through manual segmentation. The performance of
humans in recognizing ocular images using eye-brows was evaluated in [55].

Smereka et al. [135] used correlation filters for periocular region feature extraction.
Studies in [12]; [13] ; [134] used Phase Intensive Global Pattern (PIGP) and Phase
Intensive Local Pattern (PILP) for periocular feature extraction. PIGP is a global
feature descriptor that is based on variation of intensity of a pixel-neighbors with
respect to different phases [134]. PILP was extracted using a set of multi-scale filter
banks. Periocular probabilistic deformation models (PPDMs) are proposed that reduce
the image matching problem to matching local image regions and approximate the
periocular distortions by local patch level spatial translations whose relationships are
modeled by a Gaussian Markov random field [133]. Raja et al. [118] ; [121] used
Deep Sparse Filtering, that offer the convenience of unsupervised learning while
trying not to model any distributions explicitly, for periocular region representation.
Nie et al. [88], used Restricted Boltzmann Machine for periocular feature extraction.
 Moreno et al. [85] ; [86], proposed a periocular recognition model based on sparse
representations of image patches and color/geometric information that are fused into a
single representation for classification. Raghavendra and Busch [111], proposed a
periocular recognition scheme based on Maximum Response (MR) texture features.

Classification: For periocular region classification, distance metrics like Euclidean

distance [2]; [12]; [12]; [13]; [16]; [91]; [93]; [93]; [94]; [94]; [95]; [127]; [134] ; [154],
mean square error [135], city block distance [12]; [16]; [91]; [95]; [127]; [153] ; [154], chi-
square distance [4]; [5] ; [16], hamming distance [50], point-pattern matching [5]; [91]; [93];
[94]; [127]; [154] ; [154] and l1 norm [121] were used for classification. Few studies used
learning based methods such as SVM [43] ; [88], LDA [43], and Deep Neural Nets [111] ;
[118] for periocular region classification. Periocular biometrics have also been surveyed and
discussed in [7]; [8]; [89] ; [93]. Recently, Akhtar et al. [4] used combination of four local
descriptors like Multi-scale Local Binary Patterns (MLBP), Pyramid of Histogram of
Oriented Gradients (PHOG), Scale Invariant Feature Transform (SIFT) and Speeded-up
Robust Features (SURF) for personal recognition based on mobile periocular region. The final
score was estimated using weighted fusion of scores corresponding to local descriptors. The
performance of cross-sensor and cross modality periocular recognition have been evaluated
based on probabilistic deformation models [133], Neural Network [129] and local descriptors
such as SIFT, Weber Local Descriptor (WLD) and LBP [9]; [23]; [59] ; [63]. Chen and
Ferryman [24] proposed fusion of 2D and 3D periocular recognition based on LBP and
iterative closest point (ICP), respectively, at the score level. The performance of the periocular
recognition pre- and post-cataract surgery is also evaluated [65].

The performance of the periocular biometrics may be affected by facial expression change,
variability in user to device pose and angle, occlusion due to prescription or sun glasses as
well as adverse lighting conditions [15] ; [91]. Further, heavy makeup, unusually smooth skin
and physical trauma can significantly degrade the performance of periocular biometric. The
best periocular region for recognition is subjective to image size, acquisition conditions,
feature extraction and classification schemes adopted [5].

5. Proposed hardware-based acquisition set-up for ocular data

Quality of the iris image strongly influences the performance of iris recognition, which is
usually calculated in terms of image focus. Various hardware-based schemes have been
proposed to enhance the focus of the acquired eye images. For instance, Narayanswamy et
al. [87] and Boddeti and Kumar [17] explored wavefront coded imaging to improve the
overall sharpness of iridial images. Park and Kim [92] used reflection from the IR-LED
illuminator as a feedback information to set the focus and zoom. However, these above-
mentioned schemes were designed for iris recognition system operating in NIR spectrum.
Studies in [120] ; [145] proposed hardware acquisition systems for high quality and well-
focused iris images in visible spectrum discussed as follows:

Hyper-focal Acquisition System: Tankasala et al. [145] designed and implemented a

hyper-focal acquisition system for acquiring iris images in the visible spectrum. The
proposed system used a special multi-LED ring and side lights along with dSLR
Canon T2i camera. An Okii controller was programed to capture a stack of ocular
 images at successive focal lengths. The ensuing frames were fused in order to yield a
single image with higher fidelity due to increased depth of field while using a larger
aperture. The proposed setup extends the imaging depth-of-field (DOF), thereby
preempting the need for limiting the f-number to the detriment of image quality. The
Helicon focus software (Helicon Soft Ltd., Kharkov, Ukraine) was used to produce a
single hyper-focal image by fusing the sharper areas from each frame in a sequence of
8 images while adjusting for subtle movement and scaling aberrations across focal
planes.

Light-Field Camera: Raja et al. [120], proposed to capture high quality iris samples
using light-field technology to address limited depth-of-field. First generation
consumer Light-Field Camera (LFC) by Lytro Inc4 was used for iris imaging which
can provide multiple depth/focus images in a single exposure. The plenoptic/light field
cameras were constructed using a micro-lens array or a pin-hole array between the
sensor and camera lens. These micro-lenses measured the total amount of light
deposited on the sensor and direction of the incoming light. Further, by re-sorting the
measured rays of light with respect to their point of termination, a number of images
focused at different depths were obtained. The best-focused image was selected using
a wavelet energy focus measure [115] for RGB iris recognition.