Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
net/publication/314650388
Deep learning architecture for iris recognition based on optimal Gabor filters
and deep belief network
CITATIONS READS
8 580
6 authors, including:
Jinchao ji
Northeast Normal University
16 PUBLICATIONS 138 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Score level fusion scheme based on adaptive local Gabor features for face-iris-fingerprint multimodal biometric View project
All content following this page was uploaded by Han Wang on 10 April 2018.
Fei He
Ye Han
Han Wang
Jinchao Ji
Yuanning Liu
Zhiqiang Ma
Fei He, Ye Han, Han Wang, Jinchao Ji, Yuanning Liu, Zhiqiang Ma, “Deep learning
architecture for iris recognition based on optimal Gabor filters and deep belief
network,” J. Electron. Imaging 26(2), 023005 (2017),
doi: 10.1117/1.JEI.26.2.023005.
Journal of Electronic Imaging 26(2), 023005 (Mar∕Apr 2017)
Abstract. Gabor filters are widely utilized to detect iris texture information in several state-of-the-art iris recog-
nition systems. However, the proper Gabor kernels and the generative pattern of iris Gabor features need to be
predetermined in application. The traditional empirical Gabor filters and shallow iris encoding ways are incapable
of dealing with such complex variations in iris imaging including illumination, aging, deformation, and device
variations. Thereby, an adaptive Gabor filter selection strategy and deep learning architecture are presented.
We first employ particle swarm optimization approach and its binary version to define a set of data-driven Gabor
kernels for fitting the most informative filtering bands, and then capture complex pattern from the optimal Gabor
filtered coefficients by a trained deep belief network. A succession of comparative experiments validate that
our optimal Gabor filters may produce more distinctive Gabor coefficients and our iris deep representations be
more robust and stable than traditional iris Gabor codes. Furthermore, the depth and scales of the deep learning
architecture are also discussed. © 2017 SPIE and IS&T [DOI: 10.1117/1.JEI.26.2.023005]
Keywords: iris recognition; deep learning; Gabor filters; deep belief network; particle swarm optimization.
Paper 160917 received Oct. 30, 2016; accepted for publication Feb. 23, 2017; published online Mar. 10, 2017.
*Address all correspondence to: Han Wang, E-mail: wangh101@nenu.edu.cn 1017-9909/2017/$25.00 © 2017 SPIE and IS&T
Wildes et al.8 Laplacian of Gaussian filters Achieve complete description of Bring too high computation
iris texture complexity
Boles and Boashash9 Wavelet transformation Achieve complete description of Depend on the reasonable
iris texture filter selection
Monro et al.10 Discrete cosine transform Use the lower frequency Lack of time-frequency
information localization
Chang et al.11 Empirical mode decomposition Use adaptive multiresolution Lack of multiorientation
decomposition analyses
Geometric Belcher and Du12 Scale-invariant feature May be applied directly to Be sensitive to ambient
operators- transformation (SIFT) key-points iris pixels illuminations and affine
based distortions
Proença13 Structural analyses Be more stable to rotation and Rely on the quality of iris
distortion than SIFT imaging
Hybrid Tieniu et al.14 Combine iris texture, iris color, Involve more useful features Bring more external
features skin texture, and geometrical outside the irises disturbance and
asymmetry of the eyelash distribution computation complexity
Santos and Hoyle15 Combine 1-D wavelet, 2-D wavelet, Use more feasible fusion Need to adjust weighted
Gabor code, LBP, and SIFT points strategy at score level parameters
of the ocular region
Huo16 Combine Gabor filters and Apply two stages to fuse Be sensitive to the
histogram transform multiple features threshold of first stage
increase the recognition rate, because their filter banks may In addition to Gabor filters selection, this paper also
produce a better coverage in Fourier domain to extract more focuses on the iris Gabor feature generative way. In the
discriminative information. Thus, determining optimal Gabor existing systems, the most common option is to generate
filters for irises imaging from various acquisition conditions, Gabor feature vectors by applying dimension reduction or
user groups, and devices is in general problematic.25,26 Yet, binary encoding to Gabor filtered coefficients.2,7 However,
an exhaustive search of all possible parameters may not be these traditional iris representations from predefined feature
feasible, because Gabor parameter selection is a nonlinear, engineering can be hardly optimal in practice and do not
constrained programming problem with multiple continuous have the capacity to structure an accurate model, because
and discrete variables.27 Meanwhile, due to its time-frequency they are task-dependent and limited to the designer’s
transformation, it is not so easy to design a mathematical expertise.29 That restricts the superiority of Gabor filters in
objective function between Gabor kernels and iris recogni- iris texture representations. In recent years, researchers turn
tion performance. A promising solution is to utilize a two- their attention to build a flexible feature generative architec-
stage strategy, in which a set of optimized Gabor filters is ture with deep learning ability, which may be used for
searched by enrolled irises first, followed by the texture modeling the iris feature pattern in data-driven manner.
feature extraction and classification using these optimal Such deep learning methods inspired by the hierarchical
Gabor filters.25 Face recognition,24 handwritten character nature of human vision cortex may easily borrow knowledge
recognition,23 and defect detection28 have employed this from the human cognitive process.30 The deep learning archi-
two-stage strategy successfully. Inspired by the above-men- tecture can capture complex statistical patterns from Gabor
tioned two-stage approaches, in this paper, we propose an coefficients because of their uses of nonlinear operators and
iris data-driven Gabor filter optimization method based on hierarchy network.31 Neural network as a typical learning
particle swarm optimization (PSO) and its binary version architecture may generate adaptive meaningful features in
binary particle swarm optimization (BPSO), which is sup- iris recognition. However, many problems, such as local
posed to find proper Gabor kernels and avoid the unreason- minimum, long-time training, and over-fitting, would emerge
ing Gabor filter predefinition over the iris samples from when its original search strategies directly were applied
diverse image resolutions, captured devices, and user groups. to deep network architectures.32 Major breakthroughs have
been made by Hinton et al.,33 who revealed an adaptive training model parameters, it only needs a short time when
network structure using a semisupervised training strategy. predicting an unseen sample.
Benefited from this semisupervised learning strategy, deep The remainder of this article is organized as follows:
learning architecture performs well in various real-world Sec. 2 details our Gabor filter optimization method and
applications, such as face recognition,34 audio recognition,35 deep learning architecture. Section 3 validates proposed
handwritten recognition,36 and object classification.37 Among methods by a series of experimental results. Section 4 con-
existing deep learning architecture, deep belief network cludes this article.
(DBN) is a typical statistical multiple layer scheme. It is
composed by a stack of the restricted Boltzmann machines 2 Proposed Architecture
(RBM),38 each of which has a layer of input units and a
single hidden layer with hidden-to-all-visible connections. 2.1 Overview
As a generative probabilistic model, the whole DBN is The overview of our proposed iris recognition system can be
trained to achieve the maximized likelihood for a better com- found in Fig. 1. For each iris acquisition, our first task is to
pact representation of its training data. The RBM model is of separate the useful iris region from a full eye image. We com-
the nature to deal with the binary input,39 which perfectly bine Canny operator and Hough transformation to detect
matches the data type of iris Gabor codes. Furthermore, the outer boundaries of pupils.41 After that, the isolated iris
another feature of DBN is the full connected network archi- region is mapped from polar coordinates to rectangular coor-
tecture with semisupervised training strategy. Comparing dinate as a rectangular shape for normalizing the iris size.42
with other local connected networks such as convolution The normalized image will be enhanced using histogram
neuron networks (CNN), the full connected network may equalization method to improve the image quality and make
better detect potential both global and local features from the iris texture outstanding. Then, a region of interest (ROI)
iris Gabor codes. Meanwhile, the fine-tune operation in subregion, which is seldom hidden by the eyelashes and
semisupervised training strategy makes full use of category eyelids, is cut apart to provide more precise iris templates
information from the known labels and obtains the model by Yu’s ROI extraction.19
that is more suitable to address the classification problem After described above iris image preprocessing, a family
than several deep networks like stacked auto encoder of Gabor filters are applied to analyze iris texture.43 Differing
(SAE).40 Considering above described advantages, we would from existing methods using fixed empirical Gabor filters,
like to employ the DBN to further detect the hierarchical our system employs the adaptive Gabor filters determined by
representations from the binary Gabor codes of iris images. PSO and BPSO rules to fit the most informative bands of
Moreover, because the iris codes as the input of DBN are involved irises. Furthermore, most of the existing Gabor-
encoded by the optimal Gabor filters in our system, our pro- based iris descriptors depend on Gabor responses with
posed Gabor filter selection strategy may be considered as a statistic dimension reduction, while our approach puts iris
a preprocessing step to the deep learning architecture. In Gabor codes into a deep learning architecture DBN for
Table 1, we classify existed popular iris system into three detecting iris potential learning features in a data-driven
way. In such a situation, the Gabor codes of all enrollments
categories. While our proposed method may be classified
with their corresponding labels are designated to train the
into a new category called learning features. In our system,
DBN model. Finally, the convergent DBN can be used
the features directly linked to classification results are the
to implement the feature extraction and classification of
vectors at the top hidden layer, which are generated in a
test iris.
data-driven way. Such an approach is different from all men-
tioned systems in Table 1 that define iris representations
by feature engineering. Our feature generative model does 2.2 Gabor Iris Codes Generation
not need any predefined expressive pattern, but depends on Generally, an iris ORI image IðzÞ; ðx; yÞ ∈ z (z is the set of
the data involved in modeling. Even though it takes time for image pixels) is convolved with two-dimensional Gabor
kernel Gμ;ν ðzÞ to obtain Gabor coefficients as follows (for 2.3 Gabor Filter Optimization by PSO and BPSO
further details, we refer to Refs. 27 and 42): As above mentioned, we need to find appropriate Gabor fil-
ters fitting the informative bands of involving irises. In this
kkμ;ν k2 −kkμ;ν k22kzk2 h ik z σ2
i
paper, we consider that the optimal Gabor filters with the
Gμ;ν ðzÞ ¼ e 2σ e μ;ν − e− 2 : (1)
σ2
EQ-TARGET;temp:intralink-;e001;63;730
Algorithm 1 PSO- and BPSO-based Gabor filter optimization. Algorithm 2 DBN Training
- Initialize the population of particles P. - Set a random value in ½0; 1 to each connection weight w ij between
units of different layers.
- Set four-dimensional Gabor parameters ðK max ; f ; M; NÞ to each
particle p i;j by random sampling. - Set a random value in ½0; 1 to the bias b i at each hidden layer.
- Calculate the fitness of each particle p i;j according to Eq. (6), - Set the values of units on the visible layer equal to the input iris
where the Gabor filters corresponding to p i;j are used to codes.
encode z.
Unsupervised training:
- t ¼ 0 {Iterative step} for RBM in DBN do
repeat repeat
i;j
• Calculate the expectation hx i h j idata .
þ c 2 × r 2 × ½G best − p i;j ðt Þ j ¼ 1; 2; : : : ; n; (7)
• The reconstructed value of a unit x i of visible layer is
where w is the inertial weight52 in the range ½0; 1. The velocities obtained by given hidden vector h j with its bias b j
ðv i;3 ; v i;4 Þ of two discrete-value Gabor parameters M and N will be
updated by BPSO rules as X
Pðh j jx Þ ¼ σ b j þ
EQ-TARGET;temp:intralink-;e012;326;385 w ij x i : (12)
i
v i;j ðt þ 1Þ ¼ c 1 ⊗ ½P best p i;j ðt Þ
EQ-TARGET;temp:intralink-;e008;62;366
i;j
• Calculate the expectation hx i h j iconstruction .
þ c 2 ⊗ ½G best p i;j ðtÞ j ¼ 1; 2; : : : ; n; (8)
• Adjust the weight w ij according to the gradient of the log
where means “xor” operator, and þ are “and” and “or” likelihood and contrastive divergence approximation39 as
operators, respectively. c 1 and c 2 control the probability that
i;j p i;j ðt Þ and G best p i;j ðt Þ will take 1-value.
every bit of P best ∂ log Pðx Þ
The constraint of max velocity still exists in BPSO, but it just ≈ hx i h j idata − hx i hj ireconstruction ; (13)
∂w ij
EQ-TARGET;temp:intralink-;e013;326;303
• Update the position of each particle using v i;j ðt þ 1Þ. ∂ log Pðx Þ
The position components ðp i;1 ; p i;2 Þ corresponding to w ij0 ¼ ϑw ij þ η : (14)
∂w ij
EQ-TARGET;temp:intralink-;e014;326;263
(15)
• Recalculate the fitness of each particle p i;j using Eq. (6).
1
Pðv ; h; θÞ ¼ exp−E ðx ;hÞ ; (16)
• Update the best previous position of each particle Z ðθÞ
EQ-TARGET;temp:intralink-;e016;326;138
P best
i and
the best previous position of all particles G best .
where Z is to control θ to the Gaussian distribution.
until: ðt > t max Þ or G best performs above the highest expected m denotes the number of visible units and n represents
performance the amount of hidden units.
CASIA-V4-lamp OKI IRISPASS-h 640 × 480 150 10,800 1200 3000 15,000
achieves equal values between FAR and FRR in ROC curve various sizes of eyelids in normalized iris images also may
and balances the reliability and usability of an iris recogni- damage the similarity between two irises from a pattern
tion system. class. On the three iris datasets, we treat each one sample
within ROI and without ROI extraction, and then use fixed
3.2 Experimental Results Gabor filters ðK max ¼ 64; f ¼ 2; M ¼ 6; N ¼ 4Þ introduced
by literature20 to generate traditional iris codes. These Gabor
In this section, we would like to report the experiment to codes within ROI and without ROI are, respectively, used to
explain the necessity of ROI extraction first. As above conduct matching tests and obtain recognition accuracy rate
description, several noises including eyelashes may occlude
by HD. Their experimental results are shown in Fig. 2.
the effective regions of irises for feature extraction and cause
In Fig. 2, 19.7%, 21.2%, and 23.8% false nonmatches of
false nonmatching in intraclass comparisons. Moreover, the
intraclass comparisons can be prevented by the ROI extrac-
tion on CASIA-V4-interval, CASIA-V4-lamp, and JLUBR-
IRIS iris datasets, respectively. A typical example from the
two main interferences on the JLUBR-IRIS dataset is shown
in Fig. 3.
As shown in Fig. 3, the Nos. 1103 and 1141 iris images
are gathered from the same user. However, because of their
heavy eyelid and eyelash occlusions, they are classified into
two different categories in the experiments without ROI. It
means that all kinds of useful iris feature extraction should be
performed only when the ROI region may be functioned well
and the redundant eyelids and eyelashes can be excluded
from the stage. Meanwhile, the greatest improvement of
accuracy rate emerges on JLUBR-IRIS among three iris
datasets, which implies that the iris images from this dataset
contain more challenging disturbances, including eyelids and
eyelashes. The ROI extraction is thus an indispensable part in
our system. Nevertheless, there is still room for improvement
with regard to the recognition rate of the iris images with
ROI. Therefore, we take the preprocessed ROI images as
input samples in the subsequent experiments and seek further
Fig. 2 The comparisons of the iris images with ROI and without ROI improvement by applying our proposed Gabor filter optimi-
extraction.
zation and deep learning architecture.
Fig. 3 An example of false nonmatch due to eyelid and eyelash on the JLUBR-IRIS dataset.
In the next experiment, we try to implement Gabor filter In grid search, considering the computational complexity,
optimization based on PSO and BPSO with the enrollments we use a small search step in grid search and, thus, obtain
on the three datasets. For the purpose of obtaining unbiased different decimal precision parameter values between grid
Gabor filters, we conduct fivefold experiments to achieve search method and our approach in result. In Table 3, we
global optimal result with the best fitness. Meanwhile, in also list the speeds of two methods on Intel i7-6700 core
order to exhibit the effectiveness and significance of our 3.4-GHz processor. Due to the different resolutions and
Gabor optimization strategy, we also conduct other simpler amounts of enrollments among three datasets, the average
parameter search methods for comparisons. As above men- speeds of Gabor filter optimization differed on these datasets.
tioned, due to nonexistent a mathematical objective function Meanwhile, because a too small step will lead to a long-run-
between Gabor kernels and iris recognition rates, we only ning grid search, this method takes longer searching time but
may apply grid search in the comparative experiments. obtained Gabor parameters with lower precision than our
The optimization details of PSO and BPSO on three iris data- method. The results in Table 3 demonstrate that our proposed
sets are shown in Fig. 4. optimization method may achieve similar optimization
According to Fig. 4, the fitness values in each fivefold results to grid search in a comparatively short time. These
experiment can be converged around a certain value, which experimental results also indicate that the different iris data-
shows significant increases compared with the DI values with- sets lead to the adaptive Gabor filters with different Gabor
out optimization. It suggests that the Gabor codes with more parameters. In other words, the parameters of Gabor filters
discriminative and informative can be created by the optimal should be adjusted to fit the most informative bands of iris
Gabor filters. The optimized Gabor parameters from grid images in various captured device acquisition conditions and
search and our approach can be found in Table 3. participant groups. We plot the comparative ROCs using the
Fig. 4 The fitness values of PSO and BPSO on (a) the CASIA-V4-interval dataset, (b) the CASIA-V4-
Lamp dataset, and (c) the JLUBR-IRIS dataset.
Table 3 The experimental results of Gabor filter optimization on three datasets using Intel i7-6700 core 3.4-GHz processor.
Fig. 5 The comparative ROCs using the original filters and optimal Gabor filters on the three datasets.
original filters and optimal Gabor filters on the three datasets The results in Fig. 6 demonstrate that DBN needs
in Fig. 5. adequate training samples to model the architecture that pro-
As shown in Fig. 5, it explains the necessity of Gabor vides enough discriminative ability. Note that some catego-
filter optimization to various datasets. We adopt similar ries in these datasets have fairly high intraclass appearance
legends to compare the ROCs using the original and optimal variation, due to ambient illuminations, rotating acquisitions,
Gabor filters on each dataset. The EERs on three dataset per- or deformable shapes. DBN may still generate robust hier-
form to be a sharp decline from 7.32%, 14.88%, and 16.72% archical features over these variations, which reports that
to 4.16%, 8.84%, and 10.06%, respectively, which also can it may adaptively detect interesting features from the input
validate our PSO- and BPSO-based Gabor filter selection. iris codes. This superiority is more noticeable for the
Furthermore, it can be concluded from both above DIs larger-scale dataset JLUBR-IRIS. Once the DBN parameters
and ROCs that even using the optimal Gabor kernels, the have been tuned using abundant quantities of training sam-
false matching or false nonmatching still exist on three data- ples, the largest EER drop appears on the JLUBR-IRIS data-
sets. The largest drop of EERs among three datasets appears set. The lowest EERs 0.002%, 0.56%, and 1.88% on
in the experiments of the JLUBR-IRIS dataset, which CASIA-V4-interval, CASIA-V4-lamp, and JLUBR-IRIS
implies that the traditional Gabor codes may hardly handle datasets, respectively, indicate that our proposed method
well multiple challenges on larger-scale dataset. This moti- well performs under varied captured devices, various acquis-
vated us to further look for the informative iris deep repre- ition conditions, and different data scales.
sentations to replace the preliminary iris codes in iris As above-mentioned experimental results, the promising
recognition system. performance of deep hierarchical networks requires the
Next, we would like to apply the deep learning architec- usage of abundant training data. Thus, we are going to dis-
ture DBN to generate iris deep representations. Because cuss the depth of deep architecture on the adopted largest
there is no guidance about the depth setting of deep archi- scale JLUBR-IRIS dataset. We use the above-described
tecture, we stacked two RBMs to make a DBN structure
DBN with two layers of 34410-24087 hidden units as the
for each involved dataset at the beginning. Here, we sepa-
basic architecture, and insert layers with a decreasing num-
rately set 45878-32115, 21504-17203, 34410-24087 as the
ber of hidden units according to the activation probability of
number of units of the two hidden layers in DBN on the
upper layer. Therefore, we can obtain a set of DBNs with the
CASIA-V3-interval, CASIA-V3-lamp, and JLUBR-IRIS
dataset, which are on the basis of the dimension of input various numbers of hidden layers. Their numbers are from 2
Gabor codes. After training the DBNs via the greedy to 10 while the minimum amounts of hidden units of the cor-
layer-by-layer algorithm introduced in Sec. 2.4, the activa- responding DBNs are from 24,087 to 6455. We conduct iris
tion probability of the second hidden layer according to recognition experiments to this set of DBNs using the same
Eq. (11) will be lower than the first hidden layer. It can training samples. In the meantime, we also construct the
be concluded that more informative and nonlinear correla- extensive experiments using other classical deep learning
tions embedded in the Gabor codes have been detected by architecture SAE61 and CNN62 using the same experimental
the multiple layers of the DBN.37 To evaluate the advantage protocol. We equally set iris Gabor codes as input data, soft-
of the deep features, we put the EERs of the different DBNs max regression model as output layer for unbiased compar-
using varied number of training irises on three datasets isons. Specific to CNN, we use max operator as its pooling
together in Fig. 6. strategy. Figure 7 exhibits the performance of the three deep
Fig. 6 The comparative EERs of the DBNs with different portions of enrollments of involving in training
stage on the three dataset.
learning architectures with the different number of hidden of SAE is also to compress input signal to low-dimensional
layers. and sparse codes. However, the input binary Gabor codes as
From Fig. 7, we found that when the DBN fits the 34410- relatively sparse data are hard to generate further sparse rep-
24087-16542-10548-8413 hidden layers structure, the learn- resentations without information loss. In contrast, the activa-
ing performance receives peak value EER ¼ 1.84 among tion function of DBN is in the nature of better handling
three deep architectures. Putting the results of JLUBR- binary input data, and therefore, the performance of SAE
IRIS in Figs. 5 and 6 together, we can reach the conclusion cannot yet compare with our DBN architecture. CNN as a
that only when the scale of deep learning architecture and the geometric feature detector is more suitable for directly deal-
number of training samples are both sufficiently large, would ing with original image pixels, but it would not play its own
the results that the deeper learning architecture achieves the superiorities with our input iris Gabor codes. Thus, our pro-
better performance be appeared. Otherwise, it may cause posed architecture is more adaptable to further extract iris
serious over fitting. The similar attempts were carried out deep features from iris Gabor codes. Moreover, we also
by Liu et al.37 and Weston et al.63 Our conclusions are con- directly apply CNN to the original pixels of normalized ROI
sistent with their observations. It also can be concluded from images to satisfy its own suitability and then achieve better
Fig. 6 that our proposed architecture achieves outperform- results than applying it to Gabor codes. Nevertheless, these
ance among three deep learning architectures, while SAE results are not yet comparable to those of our proposed archi-
may receive similar performance with our approach. The task tecture. We believe the reason is that CNN is effective to
Fig. 7 The EERs of the different deep architectures with different hidden layers on the JLUBR-IRIS
dataset.
Table 4 The comparisons of proposed architecture and some state- best trade-off in spatial and frequency resolution, we first
of-the-art iris recognition systems on the CASIA-V4 datasets. employ Gabor filters to detect iris texture information. In
order to find proper Gabor kernels for fitting the most
CRR/% informative bands of involving irises, we propose a data-
driven Gabor filter optimization method based on PSO
Methodology CASIA-V4-interval CASIA-V4-lamp and BPSO. Next, the DBN stacked by several RBMs in semi-
unsupervised fashion is presented to further extract iris deep
Daugman 7
— 96a
features from the Gabor codes to stand for various noises in
Jan et al.64 100a 98a
real-world application. A range of comparative experiments
are conducted on the public CASIA-V4-interval and CASIA-
Ibrahim et al.65 99.9a 98.28a V4-lamp datasets as well as self-developed larger-scale
JLUBR-IRIS dataset. These experimental results indicate
Khalighi66 99.97a 99.96a that our Gabor filter selection strategy may design appropri-
ate Gabor filters for involving iris images and make an
Zhang ad Guan67 99.6b — obvious improvement of the recognition performance.
Moreover, the EERs go further down while applying DBN
Li et al.68 99.91b — to capture potential iris deep representations. The results are
superior to those of the methods using predefined Gabor fil-
Rai and Yadav69 99.91c — ters and in shallow feature generative way. In addition, the
extensive experimental results also validate that our system
Proposed method 99.998 99.904
has greater performance than recent some recent iris recog-
a
Data labeled from Ref. 66. nition approaches.
b
Data labeled from Ref. 68.
c
Data labeled from Ref. 69. Acknowledgments
In this paper, portions of the research use the CASIA-V4-
detect iris spatial features from raw image pixels, while our interval and CASIA-V4-lamp datasets collected by the
proposed deep learning architecture may extract iris potential Chinese Academy of Sciences’ Institute of Automation
patterns from the binary Gabor codes with multiscale and (CASIA). This work is partially supported by the National
multiorientation analysis. To sum up, our proposed method Natural Science Foundation of China (NSFC) under
outperforms the other two kinds of deep learning architec- Grant Nos. 61402098 and 61502093; the Science and
tures in terms of EER. Technology Development Plan of Jilin province under
Table 4 lists our results combining with the experimental Grant Nos. 20140101194JC and 20170520058JH; the
results of recent state-of-the-art iris recognition systems on Fundamental Research Funds for the Central Universities
the public CASIA-V4 dataset7,64–69 to provide unbiased com- under Grant No. 2412016KJ033; and the open project pro-
parisons. All these results are directly summarized from the gram of Key Laboratory of Symbolic Computation and
published literatures.66,68,69 Knowledge Engineering of Ministry of Education, Jilin
From Table 4, the CCR of our system has grown to University under Grant No. 93K172016K04.
99.998% and 99.904%, respectively, on the CASIA-V4-
References
interval and CASIA-V4-lamp datasets. This superior perfor-
mance benefits from the proper Gabor filter selection and the 1. A. K. Jain, A. Ross, and S. Pankanti, “Biometrics: a tool for information
security,” IEEE Trans. Inf. Forensics Secur. 1(2), 125–143 (2006).
DBN deep architecture. Some of the state-of-the-art iris 2. J. G. Daugman, “High confidence visual recognition of persons by a test
systems use wavelet transforms, including Haar, Gabor, of statistical independence,” IEEE Trans. Pattern Anal. Mach. Intell.
and contourlet, to extract iris features and then apply classi- 15(22), 1148–1161 (1993).
3. L. Shinyoung et al., “Efficient iris recognition through improvement of
fiers, such as HD, supported vector machine, and K-nearest feature vector and classifier,” Etri J. 23(2), 61–70 (2001).
neighbors to make classification.7,67,69 These wavelet-based 4. L. Ma et al., “Efficient iris recognition by characterizing key local var-
iations,” IEEE Trans. Image Process. 13(6), 739–750 (2014).
approaches generate iris representation in shallow generative 5. S. Shah and A. Ross, “Iris segmentation using geodesic active con-
way and also require proper wavelet filter selection. Thus, tours,” IEEE Trans. Inf. Forensics Secur. 4(4), 824–836 (2009).
they show less effectiveness compared with our deep 6. S. Wojciech et al., “Reliable algorithm for iris segmentation in eye
image,” Image Vis. Comput. 28(2), 231–237 (2010).
architecture with adaptive Gabor filters and iris deep repre- 7. J. Daugman, “How iris recognition works,” IEEE Trans. Circuits Syst.
sentations. Other iris systems use some geometric operators Video Technol. 14(1), 21–30 (2004).
to precisely detect texture properties in the spatial 8. R. P. Wildes et al., “A machine-vision system for iris recognition,”
Mach. Vis. Appl. 9(1), 1–8 (1996).
domain.64,65,68 Such features show more powerful discrimi- 9. W. W. Boles and B. Boashash, “A human identification technique using
native ability, but their noise sensitivity would result in mis- images of the iris and wavelet transform,” IEEE Trans. Signal Process.
46(4), 1185–1188 (1998).
alignments in certain cases. Instead, the experimental results 10. D. M. Monro, S. Rakshit, and D. Zhang, “DCT-based iris recognition,”
suggest that our deep architecture with optimal Gabor filters IEEE Trans. Pattern Anal. Mach. Intell. 29(4), 586–595 (2007).
may get improved performance comparable to these state-of- 11. C. P. Chang et al., “Using empirical mode decomposition for iris rec-
ognition,” Comput. Stand. Interfaces 31(4), 729–739 (2009).
the-art iris recognition systems. 12. C. Belcher and Y. Du, “Region-based SIFT approach to iris recogni-
tion,” Opt. Lasers Eng. 47(1), 139–147 (2009).
13. H. Proença, “An iris recognition approach through structural pattern
4 Conclusion analysis methods,” Expert Syst. 27(1), 6–16 (2010).
In this article, we present an adaptive Gabor filter selection 14. T. Tieniu et al., “Noisy iris image matching by using multiple cues,”
Pattern Recognit. Lett. 33(8), 970–977 (2012).
strategy and a DBN-based deep learning architecture for iris 15. G. Santos and E. Hoyle, “A fusion approach to unconstrained iris rec-
recognition. Because Gabor transformation can achieve the ognition,” Pattern Recognit. Lett. 33(8), 984–990 (2012).
16. G. Huo, “Secondary iris recognition method based on local energy- 47. J. Daugman, “Statistical richness of visual phase information: update on
orientation feature,” J. Electron. Imaging 24(1), 013033 (2015). recognizing persons by iris patterns,” Int. J. Comput. Vis. 45(1), 25–38
17. D. Gabor, “Theory of communication. Part 1: the analysis of informa- (2001).
tion,” J. Inst. Electr. Eng. III: Radio Commun. Eng. 93(26), 429–441 48. W. Dong, Z. Sun, and T. Tan, “Iris matching based on personalized
(1946). weight map,” IEEE Trans. Pattern Anal. Mach. Intell. 33(9), 1744–
18. F. Bianconi and A. Fernández, “Evaluation of the effects of Gabor filter 1757 (2011).
parameters on texture classification,” Pattern Recognit. 40(12), 3325– 49. Z. He et al. “Toward accurate and fast iris segmentation for iris biomet-
3335 (2007). rics,” IEEE Trans. Pattern Anal. Mach. Intell. 31(9), 1670–1684 (2009).
19. L. Yu, D. Zhang, and K. Wang, “The relative distance of key point based 50. R. C. Eberhart and J. Kennedy, “A new optimizer using particle swarm
iris recognition,” Pattern Recognit. 40(2), 423–430 (2007). theory,” in Proc. of the Sixth Int. Symp. on Micro Machine and Human
20. L. Ma, Y. Wang, and T. Tan, “Iris recognition based on multichannel Science, Vol. 1. (1995).
Gabor filtering,” in Paper presented at the Fifth Asian Conf. on 51. J. Kennedy and R. C. Eberhart, “A discrete binary version of the particle
Computer Vision, Melbourne, Victoria, Australia, Vol. 1, pp. 279– swarm algorithm,” in IEEE Int. Conf. Computational Cybernetics and
283 (2002). Simulation (1997).
21. M. Nabti and A. Bouridane, “An effective and fast iris recognition sys- 52. Y. Shi and R. Eberhart, “A modified particle swarm optimizer,” in The
tem based on a combined multi-scale feature extraction technique,” 1998 IEEE Int. Conf. on Evolutionary Computation Proc., 1998, IEEE
Pattern Recognit. 41(3), 868–879 (2008). World Congress on Computational Intelligence, pp. 69–73, IEEE
22. X. Li and M. Xie, “A novel algorithm of human iris recognition,” in (1998).
Paper presented at the IEEE Int. Symp. on Communications and 53. T. Kuremoto et al., “Time series forecasting using a deep belief
Information Technology (ISCIT), Beijing, China, Vol. 2, pp. 1231– network with restricted Boltzmann machines,” Neurocomputing 137,
1234 (2005). 47–56 (2014).
23. S. Uchimura, M. Watanabe, and Y. Hamamoto, “On the optimization of 54. X. Bai et al., “Object classification via feature fusion based marginalized
a Gabor filter-based feature extractor for handwritten character recog- kernels,” IEEE Geosci. Remote Sens. Lett. 12(1), 8–12 (2015).
nition,” Electron. Commun. Jpn. 86(12), 27–37 (2003). 55. R. J. Williams, D. E. Rumelhart, and G. E. Hinton, “Learning represen-
24. Á. Serrano et al., “Analysis of variance of Gabor filter banks parameters tation by back-propagating errors,” Nature 232, 533–536 (1986).
for optimal face recognition,” Pattern Recognit. Lett. 32(15), 1998– 56. A. Mohamed, G. E. Dahl, and G. Hinton, “Acoustic modeling using
2008 (2011). deep belief networks,” IEEE Trans. Audio Speech Lang. Process.
25. H. Guang-Hua, “Automated defect detection in textured surfaces using 20(1), 14–22 (2012).
optimal elliptical Gabor filters,” Optik 126(14), 1331–1340 (2015). 57. Specification of CASIA Iris Image Database (ver 1.0), Center for
26. T. Du-Ming, S.-K. Wu, and M.-C. Chen, “Optimal Gabor filter design Biometrics and Security Research, Chinese Academy of Sciences
for texture segmentation using stochastic optimization,” Image Vision (2005).
Comput. 19(5), 299–316 (2001). 58. H. Fei et al., “Multiple local feature representations and their fusion
based on an SVR model for iris recognition using optimized Gabor fil-
27. F. Li and K. Xu, “Optimal Gabor Kernel’s scale and orientation selec- ters,” EURASIP J. Adv. Signal Process. 2014(1), 95 (2014).
tion for face classification,” Opt. Laser Technol. 39(4), 852–857 (2007).
59. Y. Bengio, “Learning deep architectures for AI,” Found. Trends Mach.
28. J. Junfeng et al., “Supervised defect detection on textile fabrics via opti- Learn. 2(1), 1–127 (2009).
mal Gabor filter,” J. Ind. Text. 44(1), 40–57 (2014). 60. P. Y. Simard, D. Steinkraus, and J. C. Platt, “Best practices for convolu-
29. Z. Zhen et al., “Exemplar based deep discriminative and shareable fea- tional neural networks applied to visual document analysis,” in Seventh
ture learning for scene image classification,” Pattern Recognit. 48(10), Int. Conf. on Document Analysis and Recognition, pp. 958–963 (2003).
3004–3015 (2015). 61. P. Vincent, H. Larochelle, and I. Lajoie, “Stacked denoising autoen-
30. L. Jing, L. Bingyuan, and L. Hanqing, “Detection guided deconvolu- coders: learning useful representations in a deep network with a
tional network for hierarchical feature learning,” Pattern Recognit. local denoising criterion,” J. Mach. Learn. Res. 11(12), 3371–3408
48(8), 2645–2655 (2015). (2010).
31. M. A. Ranzato et al., “On deep generative models with applications to 62. Y. Lecun, L. Bottou, and Y. Bengio, “Gradient-based learning applied to
recognition,” IEEE Conf. on Computer Vision and Pattern Recognition document recognition,” Proc. IEEE 86(11), 2278–2324 (1998).
(CVPR), pp. 2857–2864, IEEE (2011). 63. J. Weston et al., “Deep learning via semi-supervised embedding,” in
32. D. Shifei et al., “Rough neural networks: a review,” J. Comput. Inf. Syst. Neural Networks: Tricks of the Trade, pp. 639–655, Springer, Berlin
7(7), 2338–2346 (2011). Heidelberg (2012).
33. G. Hinton, S. Osindero, and Y. W. Teh, “A fast learning algorithm for 64. F. Jan, I. Usman, and S. Agha, “Iris localization in frontal eye images for
deep belief nets,” Neural Comput. 18(7), 1527–1554 (2006). less constrained iris recognition systems,” Digital Signal Process. 22(6),
34. Y. Jin, J. Lu, and R. Qiuqi, “Coupled discriminative feature learning for 971–986 (2012).
heterogeneous face recognition,” IEEE Trans. Inf. Forensics Secur. 65. M. T. Ibrahim et al., “Iris localization using local histogram and other
10(3), 640–652 (2015). image statistics,” Opt. Lasers Eng. 50(5), 645–654 (2012).
35. L. Honglak et al., “Unsupervised feature learning for audio classifica- 66. S. Khalighi, “Iris recognition using robust localization and nonsub-
tion using convolutional deep belief networks,” Proc. of the 22nd Int. sampled contourlet based features,” J. Signal Process. Syst. 81(1),
Conf. on Neural Information Processing Systems (NIPS'09), pp. 1096– 111–128 (2015).
1104 (2009). 67. H. Zhang and X. Guan, “Iris recognition based on grouping KNN and
36. Y. LeCun et al., “Backpropagation applied to handwritten zip code rec- rectangle conversion,” in 2012 IEEE 3rd Int. Conf. on Software
ognition,” Neural Comput. 1(54), 541–551 (1989). Engineering and Service Science (ICSESS), pp. 131–134 (2012).
37. Y. Liu, S. Zhou, and Q. Chen, “Discriminative deep belief networks for 68. C. Li, W. Zhou, and S. Yuan, “Iris recognition based on a novel variation
visual data classification,” Pattern Recognit. 44(10), 2287–2296 (2011). of local binary pattern,” Visual Comput. 31(10), 1419–1429 (2015).
38. P. Smolensky, Information Processing in Dynamical Systems: 69. H. Rai and A. Yadav, “Iris recognition using combined support vector
Foundations of Harmony Theory, Vol. 1, pp. 194–281, University of machine and Hamming distance approach,” Expert Syst. Appl. 41(2),
Colorado, Boulder, Colorado (1986). 588–593 (2014).
39. G. E. Hinton, “Training products of experts by minimizing contrastive
divergence,” Neural Comput. 14(8), 1771–1800 (2002). Fei He received his BS, MS, and PhD degrees from the College
40. Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature
521(7553), 436–444 (2015). of Computer Science and Technology, Jilin University, China, in
41. Y. P. Huang, S.W. Luo, and E. Y. Chen, “An efficient iris recognition 2007, 2010, and 2015, respectively. He is currently a lecturer at
system,” in Proc. 2002 Int. Conf. on Machine Learning and the School of Computer Science and Information Technology,
Cybernetics, Vol. 1, pp. 450–454, IEEE (2002). Northeast Normal University, China. His research interests include
42. H. Park and K. R. Park, “Iris recognition based on score level fusion by pattern recognition, machine learning, biometrics, and software
using SVM,” Pattern Recognit. Lett. 28(15), 2019–2028 (2007). engineering.
43. S. L. Tai, “Image representation using 2D Gabor wavelets,” IEEE Trans.
Pattern Anal. Mach. Intell. 18(10), 959–971 (1996). Ye Han received her MS degree from the College of Computer
44. L. Wiskott, J. M. Fellous, and N. Kuiger, “Face recognition by elastic Science and Technology, Jilin University, in 2012. She is currently a
bunch graph matching,” IEEE Trans. Pattern Anal. Mach. Intell. 19(7), PhD candidate at the College of Computer Science and Technology,
775–779 (1997). Jilin University, China. Her research interests include bioinformatics,
45. V. Kyrki, J. K. Kamarainen, and H. Kälviäinen, “Simple Gabor feature pattern recognition, and machine learning.
space for invariant object recognition,” Pattern Recognit. Lett. 25(3),
311–318 (2004).
46. J. Bigun and J. M. H. du Buf, “N-folded symmetries by complex Han Wang received his PhD from the College of Computer Science
moments in Gabor space and their application to unsupervised texture and Technology, Jilin University, China. He is currently employed by
segmentation,” IEEE Trans. Pattern Anal. Mach. Intell. 16(1), 80–87 the Northeast Normal University as a science researcher. He focuses
(1994). on machine learning algorithm research for biological problems,
including the transmembrane protein structure prediction, biological currently a professor and supervisor of PhD candidates at Jilin
networks, and drug networks. University. He has wide research interests, mainly including bioinfor-
matics, pattern recognition, biometrics, pervasive computing, spam
Jinchao Ji received his PhD from the College of Computer Science identification, and product data management.
and Technology, Jilin University, China, in 2013. He is the author or
coauthor of more than fourteen scientific papers. His current research Zhiqiang Ma is currently the dean of the School of Computer Science
interests include cluster analysis, diffuse of influence, machine learn- and Information Technology, Northeast Normal University, China. His
ing, and educational data mining. research interests include bioinformatics, biometrics, vision under-
standing, and machine learning.
Yuanning Liu received his PhD from the College of Computer
Science and Technology, Jilin University, China, in 2004. He is