You are on page 1of 10

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
1

Iris Recognition Based on Human-Interpretable


Features
Jianxu Chen, Feng Shen, Danny Z. Chen and Patrick J. Flynn

Abstract—The iris is a stable biometric trait that has been certainty was rated from 1 to 5. The decision was made based
widely used for human recognition in various applications. on human perception of the overall texture. Analogous to
However, deployment of iris recognition in forensic applications fingerprints, one way to further promote the development of
has not been reported. A primary reason is the lack of human-
friendly techniques for iris comparison. To further promote iris recognition in law enforcement applications is to make
the use of iris recognition in forensics, the similarity between the similarity between irises interpretable so that the whole
irises should be made visualizable and interpretable. Recently, a process can be supervised and verified by human experts.
human-in-the-loop iris recognition system was developed, based Namely, the judgement should be made based on quantitative
on detecting and matching iris crypts. Building on this frame- matching of visible features in iris images.
work, we propose a new approach for detecting and matching
iris crypts automatically. Our detection method is able to capture In the literature, the study of iris recognition relevant to
iris crypts of various sizes. Our matching scheme is designed to forensics includes the recognition of iris captured in visible
handle potential topological changes in the detection of the same wavelength [8] or non-ideal conditions, such as on the move
crypt in different images. Our approach outperforms the known or at a distance [9]. There are very few results on investigating
visible-feature based iris recognition method on three different
iris recognition using human-friendly features. Known feature
datasets. Specifically, our approach achieves over 22% higher
rank one hit rate in identification, and over 51% lower equal based iris recognition methods, such as ordinal features [10],
error rate in verification. In addition, the benefit of our approach SIFT descriptors [11], and pseudo-structures [12], are neither
on multi-enrollment is experimentally demonstrated. easily interpretable nor corresponding to any physically visible
Index Terms—iris recognition, forensics, visible feature, features.
human-in-the-loop

I. I NTRODUCTION A. Iris Crypts and the Human-in-the-loop System Overview


Recently, Shen [13] developed a new human-in-the-loop iris
I RIS recognition is one of the most reliable techniques in
biometrics for human identification. The Daugman algo-
rithm [1] can achieve a false match rate of less than 1 in 200
biometric system which performs iris recognition by detecting
and matching crypts in iris images. Iris crypts are certain
billions [2]. Iris recognition techniques have been used widely relatively thin areas of iris tissue, which may appear near the
by governments, such as the Aadhaar project in India [3]. collarette or in the periphery of the iris. The visibility of iris
However, the iris is still under assessment as a biometric crypts stems from their relationship with the pigmentation and
trait in law enforcement applications. One reason that hinders structure of the iris.
the forensic deployment of iris is that iris recognition results In iris images captured under near infrared (NIR) illu-
are not easily interpretable to examiners. As discussed in [4], mination, the appearance of iris crypts has the following
“Iris Examiner Workstation” may be built analogously to the characteristics (see Figure 1):
“Tenprint Examiner Workstation”, which has been used in • The interior has a relatively homogeneous intensity that is
forensics [5]. In fingerprint recognition, a human examiner lower than that of the neighboring pixels in the exterior.
bases a decision on the number of matched minutiae on two • The boundary exhibits stronger edge evidence than either
fingerprints [6]. In contrast, common iris recognition tech- the interior or the exterior.
niques, such as Daugman’s framework [1], perform matching
on an iris code, which is the result of applying a band-
pass filter and quantizer to grayscale images. In this scenario,
the whole procedure appears as a black-box to an examiner
without the knowledge of image processing.
Fig. 1: Demonstration of the characteristics of iris crypts. Four
Experiments have shown that human examiners can perform
examples of iris crypts are shown in red boxes.
well in identity verification using iris images [7]. In [7], the
Copyright (c) 2013 IEEE. Personal use of this material is permitted. Comparing to fingerprint recognition, iris crypts may serve
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to pubs-permissions@ieee.org as the “minutiae of the iris” [14]. Thus, iris recognition was
formulated as the problem of detecting and matching iris
J. Chen, F. Shen, D. Z. Chen, and P. J. Flynn are with the Department of crypts [15]. Following the ACE-V methodology (Analysis,
Computer Science and Engineering, University of Notre Dame, Notre Dame
IN 46556 USA (email: jchen16@nd.edu, fshen1@nd.edu, dchen@nd.edu, Comparison, Evaluation, and Verification) commonly used in
flynn@nd.edu). fingerprint recognition [5], a notional human-in-the-loop iris

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
2

to the probe image, i.e., the smallest dissimilarity score (the


Evaluation step). In practice, m is a small integer, such as 10
or 20. Finally, human examiners will manually compare the
candidate images against the probe image with the human-
interpretable features labeled and the similarity between the
features in the probe image and different candidate images
(a) Raw image and Segmentation of A.
presented, so as to make the conclusion on the identity of the
probe image (the Verification step).
In verification applications, the system processes the probe
image to detect features first (Analysis). The dissimilarity
score between the probe image and the gallery image(s) of
the identity that the probe image claims to be is computed
(Comparison). The system will present the results to human
examiners, only if the dissimilarity score is lower than a
(b) Raw image and segmentation of B. threshold (Evaluation). The human examiners will inspect the
results, with the aid of detected features and similarity between
corresponding features, to accept or reject that the probe image
has the claimed identity.

B. Our Contributions
(c) Matched features in A and B. In this paper, we seek to improve the performance of the
automated iris recognition process, i.e., the first three steps
Fig. 2: (a)-(b) Unwrapped images and segmentation of two iris of the ACE-V framework. Specifically, we propose a new
images, A and B, from the same eye. (c) The matched features fully automated approach to: (1) extract human-interpretable
in A and B computed by our proposed matching algorithm. features in iris images, and (2) match the features with the
Each pair of matched features has the same color. Examples images in the database to determine the identity. Our proposed
of multiple-to-one correspondence are marked by red boxes. approach can provide reliable aid to human evaluation in a
human-in-the-loop iris recognition system.
Our new approach employs the following observations. In
recognition system would employ the following steps as a
theory, iris crypts may appear in various sizes and shapes in
scientific method [13]:
images. In practice, it is sometimes uncertain whether multiple
1) Analysis (A): Features (iris crypts) are detected on the proximal crypts are connected. Furthermore, slight differences
iris image under investigation, by a computer program in the acquired images of the same iris may alter the topology
or by trained examiners. of the detection of the same crypts from image to image. An
2) Comparison (C): A similarity (or dissimilarity) score example is shown in Figure 2. The two images in Figure 2 are
is computed by comparing detected features with the from the same eye, but acquired at different times. Examples
feature patterns in the database using a rigorous process. of the same crypts with different topologies are labeled in the
3) Evaluation (E): Preliminary conclusion is formed ac- red boxes. Yet, even though the topology of particular crypts
cording to the score(s). may vary, the overall similarity can still be determined quite
4) Verification (V): Different trained examiners do indepen- easily by a human examiner.
dent manual inspections of the preliminary conclusion, There are two main tasks in our approach: crypt detection
in order to make creditable decisions. and crypt matching. Our detection (or segmentation) algorithm
Previous experiments [16] have demonstrated that human is designed to handle multi-scale crypts. It applies a key
perception of iris crypts is consistent across different ex- morphological operation in a hierarchical manner. Human
aminers, even without full training. A recent approach [17] annotated training data is used to determine the major param-
aimed to automate the A, C, and E steps. As a consequence, eters, so that the detected crypts are similar to those obtained
an integrated human-in-the-loop iris recognition system was by human inspection.
established [13]. Below, we briefly summarize how the system In our matching algorithm, we adopt a matching model
works in the identification and verification scenarios. For a based on the Earth Mover’s Distance (EMD) [18]. This match-
completed description of the system and graphical interfaces ing model is quite general. Specifically, to handle possible
for human inspection and annotation, we refer readers to [13]. differences in crypt topology, our matching algorithm is able to
For identification, the probe image under investigation is establish correspondences between the detected crypts in two
first processed by the system to detect visible features auto- images, which can be one-to-one, one-to-multiple, multiple-to-
matically (the Analysis step). A dissimilarity score between the one, or even multiple-to-multiple matching. Additionally, due
probe image and each gallery image is computed (the Com- to different lighting conditions, there may be some false alarms
parison step). The system will retrieve m candidate images or missing detections. Not all crypts can be captured in every
from the gallery whose features have the most similar patterns image, subject to different physical conditions. Our matching

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
3

algorithm is carefully designed so that it performs robustly to resulting image is denoted by Ip . Next, f is applied in a
segmentation errors and potential appearance/disappearance of hierarchical fashion. In level 0, f (Ip , R0 ) is applied (for a
small crypts. chosen value R0 , as shown below). After that, a binary image
The segmentation algorithm may detect some blob-like BW0 is obtained by thresholding the output of f (Ip , R0 ) (with
regions not physically corresponding to iris crypts. As long threshold = 0), removing small regions, and filling holes inside
as such regions are stable, they will be accepted as human each connected component.
interpretable features, and can contribute to discriminating Suppose C = {cci , i = 1, . . . , k} is the collection of all
different irises. Our matching algorithm (Section II-B) is connected components in BW0 , which will be classified into
designed to be robust to such false positive errors. Therefore, two groups: acceptable features (AF0 ) and under-segmented
we use the term ”crypts” and ”human interpretable features” features (U F0 ). For each connected component cci ∈ C, we
interchangeably in the remainder of the paper. put cci into AF0 if
A preliminary version of the algorithm proposed herein was • size(cci ) < Sz1 , or
presented in [19]. Comparing to the prior work, the feature • Sz1 ≤ size(cci ) < Sz2 and std(cci ) < δ
detection approach has been modified here to reduce some where size(cci ) is the number of pixels in cci , Sz1 and Sz2
false positive errors. In addition, we conducted more extensive are two size parameters, std(cci ) is the standard deviation
evaluation in this paper comparing to [19]. Besides our in- of the intensity of cci ’s corresponding region in Ip , and δ
house dataset and ICE2005 [20], our approach was evaluated is the trained threshold for the standard deviation of the
on the CASIA-Iris-Interval (Version 4.0) dataset [21]. Consis- region intensity. Thresholding the standard deviation of the
tent results were obtained on three different datasets with fairly intensity of segmented regions is meant to include only those
large number of subjects and variety. In addition, the benefits regions with relatively homogenous intensity among all mid-
of multiple enrollment is demonstrated experimentally for the size features, considering the characteristics of iris crypts
human-in-the-loop iris recognition system. (see Section I). AF0 directly constitutes S0 , i.e., the selected
features in level 0. On the other hand, U F0 = C \ AF0 . A
C. Organization of the Paper binary mask, M0 , is built using all connected components in
The rest of this paper is organized as follows. Section II U F0 .
describes the approach proposed for feature detection and In level k (k > 0), f (Ip , R0 − k) is applied. BWk , AFk ,
feature matching. Section III reports experiments conducted U Fk , and Mk are obtained similarly as in level 0. But, Sk is
for evaluation and comparison with the known method [17]. the intersection of AFk and Mk−1 . In other words, all selected
Section IV concludes the paper and discusses directions for features in level k must reside within the region defined by
possible further improvement. Mk−1 . The hierarchical segmentation will terminate if U Fk =
∅ or k reaches the smallest scale T (T is pre-selected). At the
end, the final segmentation BW will be
II. M ETHODOLOGY
Our approach consists of two main steps: (a) detecting
∪tk=0 Sk , if ∃ 0 ≤ t ≤ T, s.t. U Ft = ∅

crypts and (b) matching crypts. The input is normalized iris BW =
(∪Tk=0 Sk ) ∪ U FT , if U Fk 6= ∅ for all 0 ≤ k ≤ T
images (of 64 × 512 pixels). Many algorithms and software
packages can be used for this purpose; we use the system in Figure 2 is an example of the segmentation results of
[20]. A dissimilarity score will be output for each pair of iris two iris images from the same eye. In this example, the
images under comparison. crypt topologies of the two images are different (see the red
boxes), but the overall similarity is not hard to establish. We
will explain how to quantitatively establish such similarity in
A. Feature Detection
Section II-B.
We employ a hierarchical segmentation algorithm based The parameters Sz1 , Sz2 , and δ were determined by the
on morphological reconstruction to detect crypts of different training data, a small in-house image dataset that was manually
scales. The core operation, denoted by f , is a closing-by- annotated by human examiners. In this dataset, there are 188
reconstruction top-hat transformation [17]. On a grayscale images from 94 eyes, two images for each eye. Each image
image I, f has the following formulation: was annotated by two different persons. The statistical result
f (I; r) = I c − RI c ((I ⊕ Dr )c ) (1) of the crypt sizes is summarized in Figure 4a. We select the
75th percentile as Sz1 , i.e., 148, and the upper adjacent value
Here, RA (B) is the morphological reconstruction of mask as Sz2 , i.e., 314. Then, for those crypts larger than Sz1 but
A from marker B. (⊕) is the dilation operation, while the smaller than Sz2 , the region contrast is calculated (i.e., the
structuring element is a disk of radius r, denoted by Dr . I c standard deviation of the grayscale values of the pixels within
is the complement of image I. the region). The result is presented in Figure 4b. δ is set as
The major steps of the segmentation algorithm are depicted the median value, namely, 0.06.
in Figure 3. First, the intensity of the image is rescaled to In addition, RT and R0 were determined by experiments
[0, 1]. The image background is estimated (by convolution on the training dataset. (Note: T is the index of the scale. R0
with a Gaussian kernel) and subtracted from the grayscale and RT are the radii of the structuring elements at scale 0 and
image. Then the image is smoothed by a Gaussian filter. The scale T , repectively.) R0 is the scale that is able to capture

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
4

Fig. 3: An overview of our segmentation algorithm. An input image is first pre-processed by background subtraction and
smoothing. On the resulting image, a morphological operation is applied in a hierarchical fashion. At level k, BWk contains
all features detected by the corresponding morphological operation. For k > 0, the selected features Sk = AFk ∩ Mk−1 , and
the region mask Mk = U Fk . AFk and U Fk are acceptable features and under-segmented features in level k, respectively. The
final segmentation collects all selected features.

(a) (b)
Fig. 4: The box-plot of the statistical results obtained from
human annotated images: (a) the sizes of crypts (in pixels), Fig. 5: Comparison of the segmentation results of an iris
and (b) the contrast of crypts of sizes in [Sz1 , Sz2 ). image using the preliminary method in [19] and our new
approach. (A) The raw image; (B) the image after background
subtraction; (C) the segmentation using [19]; (D) the current
the largest crypts in the samples. Any scale smaller than RT segmentation result. Several false positive errors by the pre-
will be able to detect only tiny features, which are usually not liminary method are indicated by red arrows. The green box
crypts. (We use RT = 3 and R0 = 8.) shows a feature not completely detected by the preliminary
Comparing our preliminary work [19], the new segmen- method.
tation scheme can obtain finer features, namely reducing
false positive errors and detecting more subtle features. One
example is shown in Figure 5 to compare the results of [19] dissimilarity between P and Q (a lower score means a higher
and the new segmentation scheme. similarity).
First, a simple registration [17] is applied to compensate for
the possible shift between the two images when unwrapping
B. Feature Matching the iris annulus. Then, to reduce the computation overhead,
The objective of this step is to measure the similar- a pre-check is performed so that obviously unmatched iris
ity/dissimilarity between two iris images based on the de- images can be discarded. P and Q are considered as a non-
tected features. Suppose P = {p1 , p2 , . . . , pn } and Q = match, i.e., the dissimilarity score is equal to 1, if
{q1 , q2 , . . . , qm } are the detected visible features (i.e., con- •
kP ∩Qk
< σ1 , or
max{kP k,kQk}
nected regions) in two images under comparison. A score
min{kP k,kQk}
ranging from 0 to 1 will be computed to determine the • max{kP k,kQk} < σ2

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
5

where σ1 and σ2 are pre-determined thresholds. We use σ1 = derivatives, we know that F1 (t) amplifies the discriminability
0.25 and σ2 = 0.5 in our approach. Intuitively, it means that of the symmetric difference in the range of [0.5 − h, 0.5 + h],
P and Q have to overlap more than 25% or their sizes differ and weakens the discriminability in the range of [0, 0.5 − h)
less than 50%. and (0.5 + h, 1], where h = 0.2063. F2 (t) has a similar
Next, the correspondence between the features in P and property.
in Q is computed using the Earth Mover’s Distance (EMD) 2) EMD Matching Model: EMD is a similarity measure
matching model [18] (see the details in Section II-B2 for this). for comparing multi-dimensional distributions. In computer
The output of the EMD matching model is a collection of vision, EMD was first introduced in [22] for image retrieval.
pairs of matched regions. Each match pair may correspond In [18], an EMD based matching model was proposed for
to a match between two features, or two sets of multiple establishing correspondence between bacteria in consecutive
features. Suppose k matched pairs are found between P and image frames of time-lapse videos. Some major advantages of
Q, namely, P¯1 = {p11 , . . . , p1n1 } ∼ Q̄1 = {q11 , . . . , qm 1
1
}, this EMD matching model include the capability of multiple-
. . ., P¯k = {pk1 , . . . , pknk } ∼ Q¯k = {q1k , . . . , qm
k
k
}. To take to-multiple matching of objects and the robustness to dealing
the potential iris deformation and movement into account, the with various segmentation errors [18].
dissimilarity between each pair of P̄i and Q̄i is computed as Below, we briefly summarize the EMD based match-
the minimum of {Sim(P¯i0 , Q̄i ) | P¯i0 is transformed from P̄i by ing model [18]. The input consists of two signatures,
shifting up to ±h pixels in the horizontal direction or vertical P = {(p0 , wp0 ), (p1 , wp1 ), . . . , (pn , wpn )} and Q =
direction}, where Sim(∗, ∗) is the feature dissimilarity that {(q0 , wq0 ), (q1 , wq1 ), . . . , (qm , wqm )}. In the context of visible
will be defined in Section II-B1. Next, the k matched pairs are feature matching here, pi (1 ≤ i ≤ n) and qj (1 ≤ j ≤ m)
ranked. The final dissimilarity score is the weighted arithmetic are detected features in two iris images, while wpi and wqj
mean of the top r% pairs, while the weight of the i-th rank are the sizes of the features pi and qj , respectively; p0 and
pair is computed by the following formula: q0 are auxiliary variables, both of size infinity (needed by the
spi + sqi 1 processing of the model).
∗ min(spi ,sqi )
(2) A key component of EMD is the ground distance between
2 1 + exp(−10 ∗ + 2)
max(spi ,sqi ) pi and qj , denoted by Dij . Intuitively, if one views each pi as
where spi (resp., sqi ) is the size of P̄i (resp., Q̄i ). Basically, a pile of dirt (of an amount wpi ) and each qj as a hole (of a
if a match pair is more reliable, then it will have a larger volume wqj ), then Dij represents the cost for moving one unit
contribution (i.e., a larger weight) in the final score. Here, the of dirt from pi to qj . Let fij denote the mass of flow from
reliability of a match pair is evaluated in two aspects. The pi to qj . Then, EMD measures the smallest average cost for
first term in Formula (2) measures the average size of the mass distribution from P to Q under certain linear constraints.
matched pair. The second term in Formula (2) is to assess the Precisely, it has the following definition.
size difference between the matched pair (rescaled to (0, 1)). Minimize the objective function
Finally, the smaller the final dissimilarity score is, the more
P P
Dij fij
similar the two images are. EM D(P, Q) =
0≤i≤n 0≤j≤m
P P (4)
1) Feature Dissimilarity: To measure the dissimilarity be- fij
tween two features or two sets of features, A and B, the 0≤i≤n 0≤j≤m

following equation is used. subject to


X
Sim(A, B) = α∗F1 (Sym(A, B))+(1−α)∗F2 (Haus(A, B)) fij ≤ wpi , ∀ 1 ≤ i ≤ n (5a)
(3) 0≤j≤m
Sym(A, B) is the symmetric difference between A and B. X
fij ≤ wqj , ∀ 1 ≤ j ≤ m (5b)
Haus(A, B) is the Hausdorff distance between A and B. α 0≤i≤n
is a weight parameter (α = 0.5 in our implementation). F1 (∗) X X X X
and F2 (∗) are two sigmoid functions to rescale the symmetric fij = min( w pi , wqj ) (5c)
difference and the Hausdorff distance: 0≤i≤n 0≤j≤m 1≤i≤n 1≤j≤m
1
• F1 (t) = 1+exp(−10t+5) fij ≥ 0 , ∀ 0 ≤ i ≤ n , 0 ≤ j ≤ m (5d)
1
• F2 (t) = 1+exp(−t+3) In our problem, Dij (with both i > 0 and j > 0) is
The proposed dissimilarity measure is designed to evaluate the dissimilarity between the features pi and qj , namely,
the shape similarity in terms of two different aspects. In Sim(pi , qj ), as described in Section II-B1. For either i > 0
general, the systemic difference is to evaluate how much two or j > 0 but not both, Dij = 1, which is an upper bound of
regions overlap. On the other hand, the Hausdorff distance the function Sim(∗, ∗). This is the cost for a special “feature”
tends to focus on how far each point in one region locates that matches to nothing. D00 is set as +∞ in order to avoid
away from the other region, no matter how much the two correspondence between the two auxiliary variables p0 and q0 .
regions overlap. To combine these two shape similarity mea- In essence, EMD can be solved as a transportation problem
sures together, sigmoid functions are used to rescale them in polynomial time [23]. From the perspective of a trans-
into the same range, i.e., (0, 1). Another advantage of the portation problem, pi is a source and qj is a destination, and
sigmoid functions is to emphasize the similarity difference in their weights are respectively the amount available at a source
a particular range. After computing the first and second order and the amount demanded by a destination. Then, constraint

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
6

were 2639 images from 395 eyes, each with multiple images
enrolled. 395 images were randomly selected, each from a
unique eye, to form the gallery dataset. The probe dataset con-
sisted of the remaining 2244 images. Due to the randomness
of the gallery set and probe set partition, the experiments in
all the three datasets were repeated ten times in order to obtain
statistically valid results.
The performance of our method on the dataset mixing all the
Fig. 6: Illustrating the EMD based matching model: Given three datasets (i.e., with 7775 probe images and 1340 gallery
P = {p1 , . . . , p5 } and Q = {q1 , . . . , q6 } as the features images/subjects) will also be reported. All original NIR images
detected in two iris images, a bipartite graph is built in the were pre-processed and unwrapped into images of 64 × 512
EMD matching model (some of the edges are omitted for clear pixels by the irisBEE software [20]. Our proposed automated
visualization). p0 and q0 are auxiliary variables. Three matched approach was implemented and tested in Matlab, with the
pairs are obtained by solving the model. Here, q3 is a false unwrapped images as input. Our approach was compared with
positive detection that matches to nothing. the method of Shen and Flynn [17].

B. Identification
(5a) (resp., (5b)) restricts that each source (resp., destination)
can only send (resp., receive) no more than the available In the experiments of human identification, each probe
amount (resp., the demanding amount); constraint (5c) means image was compared against all gallery images to determine
that we have to either empty all available supplies at the the identity of the probe image. The top m (say 10) candi-
sources or satisfy all demands at the destinations; constraint dates with the smallest dissimilarity scores were presented to
(5d) ensures that the flows can only move from the sources to human examiners for further inspection. This was a closed-
the destinations. set comparison. Namely, it was known that at least one image
By solving the above optimization problem, it is easy to from the same subject had been enrolled in the gallery set.
interpret the resulted correspondence between the features in Before selecting the candidates, a pre-check was imposed.
P and in Q when {fij } achieving optimality. Specifically, a Suppose the k-th gallery image has nk matched features with
large amount of flow between pi and qj indicates a strong the probe image. Then, any gallery image with less than
correspondence between two crypts pi and qj . Finally, the 0.5 × M ax{nk , for all k} matched features will be considered
output is a collection of matched features. A matching example as non-match. Among the remaining gallery images, the m
is shown in Figure 6. gallery images with the smallest dissimilarity scores were
output as candidates.
III. E XPERIMENTS AND R ESULTS The cumulative match characteristics (CMC) was adopted
as the metric. Generally, the accuracy at rank m represents the
A. Datasets and Software probability that the correct subject is in the top m candidates.
We conducted experiments on three datasets, our in-house In forensic applications, we hope to return a small set of
dataset [17], ICE2005 [20], and CASIA-Iris-Interval (v4) [21], candidates to professional examiners for further inspection,
in order to evaluate our proposed iris recognition approach in while the correct subject has a high probability to be within
both the identification and verification scenarios. Our in-house the selected candidates.
dataset [17] contained 3505 images from 701 eyes, five images For the in-house dataset, the results of our proposed ap-
for each eye. In the experiments, one image of each eye was proach and the method of Shen and Flynn [17] were plotted
randomly selected as the gallery image, while the other four in Figure 8. It was demonstrated that our approach achieved
images of the same eye were used as probe images. Thus, the at least 22% higher rank one hit rate than [17]. For ICE2005,
probe set contained 2804 images, and there were 701 images the comparison was illustrated in Figure 9, showing that the
in the gallery set. The in-house dataset will be released at rank one hit rate of our approach was at least 58% higher than
http://www.nd.edu/∼cvrl. that of [17]. Figure 10 shows the performance on the CASIA-
In ICE2005 [20], there were 2953 images. Two images Iris-Interval dataset; our approach has 56% higher rank one hit
were rejected by irisBEE in the pre-processing stage due to rate than [17]. Furthermore, the 95% confidence intervals of
off-angle iris. ICE2005 contained images from 244 different the results on the dataset mixing all the three datasets and
eyes, 175 eyes with multiple images and 69 eyes with only on each individual dataset are shown in Figure 11. In the
one image enrolled. Thus, the gallery dataset consisted of results of our approach, on all datasets if we select the top
244 images, each randomly selected for a unique eye. The 10 candidates for further inspection, the probability that the
remaining 2707 images formed the probe set. true image is returned was higher than 95%.
The CASIA-Iris-Interval dataset (Version 4.0) [21] was The errors incurred by our approach were mainly due to
collected by the Chinese Academy of Sciences’ Institute of blurry images, high occlusion by eyelids or eyelashes, and
Automation (CASIA), and captured by a novel self-developed large deformation caused by off-angle iris. Examples of such
camera. The images presented very detailed textures, which errors are shown in Figure 7. Thus, additional pre-check at
were good for visible feature detection. In this dataset, there image acquisition to remove low quality images or advanced

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
7

Fig. 9: The CMC curves of the identification results of our


Fig. 7: Examples of iris images difficult to recognize in our proposed approach and the method of Shen and Flynn [17] on
experiments. (A1-A2) Two different images from the same iris. ICE2005.
The red arrows show several crypts visible in (A2), but hard
to detect in (A1), due to blurry artifacts. More examples failed
to recognize are (B) out-of-focus, (C) deformed iris sheet, and
(D) iris with heavy occlusion.

Fig. 10: The CMC curves of the identification results of our


proposed approach and the method of Shen and Flynn [17] on
CASIA-Iris-Interval.
Fig. 8: The CMC curves of the identification results of our
proposed approach and the method of Shen and Flynn [17] on Figures 15, 16, and 17). Also, the 95% confidence intervals
the in-house dataset. of the results on the dataset mixing all the three datasets
and on each individual dataset are demonstrated in Figure 18.
algorithms to enhance image quality will be helpful for further Meanwhile, we calculated the Equal Error Rate (EER), which
improvement. is defined as the common value when the false acceptance rate
is the same as the false rejection rate. In general, the smaller
C. Verification the EER is, the more accurate the method is. The results are
summarized in Table I. In short, our approach reduced the EER
In the human verification experiments, our objective was to over [17] by almost 51% on the in-house dataset, by 84% on
determine whether two images were from the same subject, ICE2005, and by 91% on CASIA-Iris-Interval.
based barely on the dissimilarity score. First, the impostor
(non-match) distribution and the authentic (match) distribution
of the results were analyzed. The comparisons between our D. Multi-Enrollment
proposed approach and the method of Shen and Flynn [17] In general, the recognition performance can be improved by
on the three datasets were plotted in Figures 12, 13, and multiple enrollment. Namely, multiple images of the same eye
14, respectively. It was evident that our approach showed are enrolled in the database and labeled as the same identity.
significant improvement in terms of discrimination over [17]. Our proposed approach can provide more reliable aid to human
It is worth mentioning that there are sudden hikes (near x = 1)
in the non-match distribution of our method. This is due In-House ICE2005 CASIA
to assigning the dissimilarity as one if two images under Ours 0.020 0.0358 0.0139
comparison cannot pass the pre-check. [17] 0.041 0.223 0.153
Moreover, the Receiver Operating Characteristic (ROC)
curve was used for further evaluation on the datasets (see TABLE I: The equal error rates of the verification experiments.

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
8

Fig. 14: The match (authentic) and non-match (imposter)


distributions of our proposed approach and the method of Shen
Fig. 11: The CMC curves with error bars of the identification and Flynn [17] on CASIA-Iris-Interval.
results of our proposed approach on CASIA-Iris-Interval, in-
house data, ICE2005, and the one mixing these three datasets.

Fig. 12: The match (authentic) and non-match (imposter) Fig. 15: The ROC curves of the verification results of our
distributions of our proposed approach and the method of Shen proposed approach and the method of Shen and Flynn [17] on
and Flynn [17] on the in-house dataset. the in-house dataset.

examiners in the human-in-the-loop iris recognition system detection and matching can provide much more reliable aid
by taking multiple enrollment. Experiments were conducted for human inspection.
on our in-house dataset to demonstrate the effect of multiple
enrollment. There were five images for each eye in the dataset.
IV. C ONCLUSIONS AND F UTURE W ORK
For each eye, one image was randomly selected as probe
image, and different numbers of images (1, 2, 3, or 4) were en- In this paper, we present a new approach for detecting and
rolled as gallery images. The results are summarized in Figure matching iris crypts for the human-in-the-loop iris biometric
19, showing that multiple enrollment can greatly improve the system. Our proposed approach produces promising results
performance. That is, in this scenario, the automated feature

Fig. 13: The match (authentic) and non-match (imposter) Fig. 16: The ROC curves of the verification results of our
distributions of our proposed approach and the method of Shen proposed approach and the method of Shen and Flynn [17] on
and Flynn [17] on ICE2005. ICE2005.

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
9

terms of subject verification.


Note that the three datasets under evaluation were collected
using different facilities among different population groups.
Also, the parameters used in our approach were trained on
another small set of homemade data. The generality and
effectiveness of our approach on diverse image data can be
demonstrated. Furthermore, as far as we know, this work
is so far the only evaluation of a human-interpretable iris
feature matching approach using public datasets (ICE2005 and
CASIA-Iris-Interval), which offers a direct comparison with
traditional approaches such as Daugman’s framework [1].
To further increase the reliability of the human-in-the-
loop iris biometric system, incorporating a quality measure
Fig. 17: The ROC curves of the verification results of our for images enrolled in the system would be beneficial. This
proposed approach and the method of Shen and Flynn [17] on would allow to evaluate whether the quality of each acquired
CASIA-Iris-Interval. image is good enough for visual feature matching. Based
on our observations and trial studies, our approach is robust
with respect to certain common factors, such as interlacing
or moderate blurring. But, it may still be affected by other
factors. Similar to the conventional approaches, high dilation
in pupil, off-angle iris, and severe blurring would be some of
the important factors. In addition, heavy occlusion and bad
illumination would have more severe effect on our approach
than the traditional iris code. Specifically, huge occlusion may
significantly reduce the number of visual features that can
be used for matching. Poor illumination may result in low
contrast so that less features can be detected than under normal
illumination. Nevertheless, our approach under the human-
in-the-loop iris recognition framework exhibits a promising
application of the iris as a biometric trait in forensics.
Fig. 18: The ROC curves with error bars of the verification
results of our proposed approach on CASIA-Iris-Interval, in- ACKNOWLEDGMENT
house data, ICE2005, and the one mixing these three datasets.
D.Z. Chen’s research was supported in part by NSF under
Grant CCF-1217906.
on all the three tested datasets, in-house dataset, ICE2005,
R EFERENCES
and CASIA-Iris-Interval. Comparing to the known method,
our approach improves the iris recognition performance by at [1] J. Daugman, “How iris recognition works,” IEEE Trans. Circuits and
Systems for Video Technology, vol. 14, no. 1, pp. 21–30, 2004.
least 22% on the rank one hit rate in the context of human [2] ——, “Probing the uniqueness and randomness of IrisCodes: Results
identification and by at least 51% on the equal error rate in from 200 billion iris pair comparisons,” Proc. IEEE, vol. 94, no. 11, pp.
1927–1935, 2006.
[3] Unique Identification Authority of India. [Online]. Available:
http://uidai.gov.in.
[4] K. R. Nobel, “The state of the art in algorithms, fast identification
solutions and forensic applications,” MorphTrush USA, Tech. Rep.,
January 2013. [Online]. Available: http://www.planetbiometrics.com/
article-details/i/1446/
[5] P. E. Peterson, C. B. Dreyfus, M. R. Gische, M. Hollars, M. A. Roberts,
R. M. Ruth, H. M. Webster, and G. L. Soltis, “Latent prints: A per-
spective on the state of the science,” Forensic Science Communications,
vol. 11, no. 4, 2009.
[6] C. Champod, “Edmond Locard — numerical standards and ”probable”
identifications,” Journal of Forensic Identification, vol. 45, no. 2, pp.
136–163, 1995.
[7] K. McGinn, S. Tarin, and K. W. Bowyer, “Identity verification using
iris images: Performance of human examiners,” in IEEE Conf. on
Biometrics: Theory, Applications and Systems (BTAS), 2013, pp. 1–6.
[8] H. Proenca, “Iris recognition: On the segmentation of degraded images
acquired in the visible wavelength,” IEEE Trans on Pattern Analysis and
Machine Intelligence, vol. 32, no. 8, pp. 1502–1516, 2010.
Fig. 19: The ROC curves of the verification results of our [9] H. Proenca, S. Filipe, R. Santos, J. Oliveira, and L. Alexandre, “The
proposed approach on the in-house dataset, when different ubiris. v2: A database of visible wavelength iris images captured on-the-
move and at-a-distance,” IEEE Trans on Pattern Analysis and Machine
numbers of images are enrolled for the same eye. Intelligence, vol. 32, no. 8, pp. 1529–1535, 2010.

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2016.2535901, IEEE
Transactions on Information Forensics and Security
10

[10] Z. Sun, L. Wang, and T. Tan, “Ordinal feature selection for iris and
palmprint recognition,” IEEE Trans. Image Processing, vol. 23, no. 9,
pp. 3922–3934, 2014.
[11] M. Sunder and A. Ross, “Iris image retrieval based on macro-features,”
in Proc. International Conf. Pattern Recognition,, 2010, pp. 1318–1321.
[12] J. D. Mira and J. Mayer, “Image feature extraction for application of
biometric identification of iris — a morphological approach,” in Proc. Feng Shen received the B.S. degree in Electrical
Brazilian Symp. Computer Graphics and Image Processing, 2003, pp. Engineering (2006) and the M.S. degree in Pattern
391–398. Recognition (2009) from Shanghai Jiao Tong Uni-
[13] F. Shen, “A visually interpretable iris recognition system with crypt versity. He received the Ph.D. degree in Computer
features,” Ph.D. dissertation, University of Notre Dame, 2014. Science and Engineering from University of Notre
[14] F. Shen and P. J. Flynn, “Using crypts as iris minutiae,” in Proc. SPIE Dame in 2015. He worked with advisor Dr. Patrick
8712, Biometric and Surveillance Technology for Human and Activity Flynn in investigating the use of visible features
Identification X, 2013, p. 87120B. in iris biometrics. His research interests include
[15] ——, “Iris matching by crypts and anti-crypts,” in Proc. IEEE Conf. biometrics, computer vision and machine learning.
Technologies for Homeland Security, 2012, pp. 208–213.
[16] ——, “Are iris crypts useful in identity recognition?” in Proc. IEEE
International Conf. Biometrics: Theory, Applications and Systems, 2013,
pp. 1–6.
[17] ——, “Iris crypts: Multi-scale detection and shape-based matching,” in
Proc. IEEE Winter Conf. Applications of Computer Vision, 2014, pp. Danny Ziyi Chen is a professor in the Department
977–983. of Computer Science and Engineering at the Uni-
[18] J. Chen, C. W. Harvey, M. S. Alber, and D. Z. Chen, “A matching model versity of Notre Dame. His main research interests
based on earth mover’s distance for tracking Myxococcus Xanthus,” in are in the areas of algorithm design, analysis, and
Proc. Medical Image Computing and Computer-Assisted Intervention, implementation, computational geometry, computa-
2014, pp. 113–120. tional biomedicine, biomedical imaging, data min-
[19] J. Chen, F. Shen, D. Z. Chen, and P. J. Flynn, “Iris recognition based ing, parallel and distributed computing, robotics, and
on human-interpretable features,” in Proc. IEEE International Conf. on VLSI design. He has published over 310 journal and
Identity, Security and Behavior Analysis (ISBA), 2015, pp. 1–6. conference papers and holds five US patents in these
[20] P. J. Phillips, K. W. Bowyer, P. J. Flynn, X. Liu, and W. T. Scruggs, areas. He is a fellow of the IEEE and a distinguished
“The iris challenge evaluation 2005,” in Proc. IEEE International Conf. scientist of ACM.
Biometrics: Theory, Applications and Systems., 2008, pp. 1–8.
[21] CASIA Iris Image Database. [Online]. Available:
http://biometrics.idealtest.org/.
[22] Y. Rubner, C. Tomasi, and L. J. Guibas, “A metric for distributions
with applications to image databases,” in Proc. IEEE International Conf. Patrick J. Flynn is the Duda Family Professor of
Computer Vision, 1998, pp. 59–66. Engineering, at the University of Notre Dame. He
[23] F. S. Hillier and G. J. Lieberman, Introduction to Operations Research, received the Ph.D. in Computer Science in 1990
7th ed. McGraw-Hill, 2001. from Michigan State University. He has held faculty
positions at Notre Dame, Washington State Univer-
Jianxu Chen received his B.S. degree in Informa- sity, and The Ohio State University. His research
tion and Computational Science from the University interests include computer vision, biometrics, and
of Science and Technology of China, in 2011. He image processing. Dr. Flynn is an IEEE Fellow, an
is currently pursuing the Ph.D. degree from the IAPR Fellow, and an ACM Distinguished Scientist.
University of Notre Dame under the supervision He is the Editor-in-Chief of the IEEE Biometrics
of Dr. Danny Chen. His research interests include Compendium, and is a past Associate Editor-in-
computer vision, machine learning, image process- Chief of IEEE Trans. on PAMI, and a past Associate Editor of IEEE TIFS,
ing and biomedical imaging. IEEE TIP, IEEE TPAMI, Pattern Recognition, and Pattern Recognition Letters.

1556-6013 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like