Learning Domain Specific Features using Convolutional Autoencoder:
A Vein Authentication Case Study using Siamese Triplet Loss Network
Manish Agnihotri
1
, Aditya Rathod
1
, Daksh Thapar
2
, Gaurav Jaswal
2
,
Kamlesh Tiwari
3
and Aditya Nigam
2
1
Department of Information & Communication Technology, Manipal Institute of Technology Manipal, India
2
School of Computing and Electrical Engineering, Indian Institute of Technology Mandi, India
3
Department of Computer Science and Information Systems, Birla Institute of Technology and Science Pilani, India
Keywords:
Siamese Network, Vascular Biometrics.
Abstract:
Recently, deep hierarchically learned models (such as CNN) have achieved superior performance in various
computer vision tasks but limited attention has been paid to biometrics till now. This is major because of the
number of samples available in biometrics are limited and are not enough to train CNN efficiently. However,
deep learning often requires a lot of training data because of the huge number of parameters to be tuned by the
learning algorithm. How about designing an end-to-end deep learning network to match the biometric features
when the number of training samples is limited? To address this problem, we propose a new way to design
an end-to-end deep neural network that works in two major steps: first an auto-encoder has been trained for
learning domain specific features followed by a Siamese network trained via. triplet loss function for matching.
A publicly available vein image data set has been utilized as a case study to justify our proposal. We observed
that transformations learned from such a network provide domain specific and most discriminative vascular
features. Subsequently, the corresponding traits are matched using multimodal pipelined end-to-end network
in which the convolutional layers are pre-trained in an unsupervised fashion as an autoencoder. Thorough
experimental studies suggest that the proposed framework consistently outperforms several state-of-the-art
vein recognition approaches.
1 INTRODUCTION
Authentication Systems based on biometrics have
been increasingly used in a variety of security appli-
cations such as banking, immigration control, foren-
sics and so on. Several biometric traits like iris, fin-
gerprint, ear, face etc. have been studied for reli-
able security solutions. However, each trait has its
own specific challenges that limit its usages in real
time scenarios. Moreover, using a unimodal sys-
tem makes it vulnerable to presentation attacks (Patil
et al., 2016). As the multimodal systems blend the
information from multiple biometric traits, it not only
increases the performance of the system by comple-
menting the lapses of each individual traits, it also
makes the system secure from attacks like presen-
tation attacks. Moreover, they also provide better
population coverage, which are essential for perform-
ing recognition on large databases. Apart from such
benefits, multiple sensors are required to capture the
multi-biometric samples, but that increase the overall
cost and require a higher degree of user cooperation
(Kimura et al., 2015). However, the above-mentioned
shortcomings can be improved if the biometric char-
acteristics lie close to each other. For instance, the
acquisition of frontal hand region offers this opportu-
nity to capture the biometric characteristics together
from finger and palm regions. Moreover, uniqueness
of the blood vessel networks among the individuals
and challenges to its reproduction make vein patterns
a strong biometric identifier.
1.1 Related Work
Multimodal biometric systems involve the integration
of evidence of information from multiple biometric
traits to achieve a performance superior to any of the
individual modalities. However, the choice of bio-
metric traits plays a key role in the usability of the
multimodal biometric systems. More specifically, re-
778
Agnihotri, M., Rathod, A., Thapar, D., Jaswal, G., Tiwari, K. and Nigam, A.
Learning Domain Specific Features using Convolutional Autoencoder: A Vein Authentication Case Study using Siamese Triplet Loss Network.
DOI: 10.5220/0007568007780785
In Proceedings of the 8th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2019), pages 778-785
ISBN: 978-989-758-351-3
Copyright
c
 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

search in vascular biometrics has become very pop-
ular. Besides being unique, the subcutaneous vein
structures have the added advantage of lying under-
neath the skin surface. This makes visibility to the
eyes or general purpose cameras difficult and hence,
this limits the ease of spoofing if not averting it com-
pletely (Patil et al., 2016). For instance, the au-
thors in (Yang and Zhang, 2012) uses both finger-
print and finger-vein modalities as they are both ex-
tracted from the finger region, making the multimodal
biometric collection process convenient. Similar re-
searches have been carried on for palmprint and palm-
vein (Wang et al., 2008), finger-vein and finger dorsal
texture (Yang et al., 2014), hand dorsal and palm vein
patterns (Ramalho et al., 2011), finger knuckle and
finger vein (Veluchamy and Karlmarx, 2016) and so
on. Although these approaches require lesser effort
on the user’s part, nevertheless leads to the employ-
ment of more complex hardware. The other studies in
vascular biometrics explore the line based and curva-
ture based information in the vessel structures present
in the biometric samples (Zhou and Kumar, 2010).
Also, it has been noted that much attention has been
paid to the palm-vein and finger-vein modalities in-
dividually, but very few researchers attempted to ad-
dress the problem of presentation attack for vascular
technology (Choi et al., 2009). However, much recent
work has been focused on employing the deep learn-
ing techniques in various domains and the field of bio-
metrics is not an exception. These works have shown
that the deep learning based features perform better
than the handcrafted features for face (Taigman et al.,
2014) and finger-vein (Qin and El-Yacoubi, 2017)
verification systems. In (Xie and Kumar, 2017), au-
thors demonstrated a new deep learning approach for
the finger vein recognition using the CNN and super-
vised discrete hashing. In a very recent study (Fang
et al., 2018), authors proposed a two-channel CNN
network that has only three convolution layers for fin-
ger vein verification.
(a) (b)
Figure 1: Vein ROI Extraction (Bhilare et al., 2018): (a)
Palm-vein ROI, (b) Finger Vein ROI.
1.2 Challenges and Contribution
Challenges: The vein patterns either collected from
dorsal or palmer side of the hand, provide very vast
textural information. The acquisition procedures are
convenient and hygienic than other methods. But,
substantial changes in hand positioning during con-
tactless acquisition, make this task very difficult to
achieve a satisfactory ROI segmentation since that af-
fects the overall system performance. The other open
issues in vein verification are the lack of robustness
against image quality degradation, illumination, and
point-wise matching. A few existing image trans-
formation techniques (such as LBP (Ahonen et al.,
2006), BOP(Jaswal et al., 2017), TCM (Umer et al.,
2016), GOP(Nigam et al., 2016), IRT(Cummings
et al., 2011) etc.) are well proposed in literature that
create useful representation of image data and helps to
improve the matching task (Jaswal et al., 2017). But
no work has been proposed yet that encodes the im-
age feature through a Deep Learning model. Earlier
works in computer vision and big data have focused
their attention on object detection, feature extraction
and matching using deep learning models. There-
fore, efforts have to be made to bridge the gap be-
tween deep learning and biometric recognition. To
best of our knowledge, this is the first attempt in
which, a convolutional autoencoder has been trained
to learn the Texture Code Matrix (TCM) and Image
Ray Transform (IRT) based encoding schemes to ob-
tain the deep domain specific features for palm vein
and finger vein modalities. By doing such network
learned transformations, we use network learned fea-
tures for matching and achieve speedy computations
that surpasses the conventional TCM and IRT based
matching.
Contribution: The major contribution of this work
includes four folds. (i) A Deep Learning based vein
recognition framework has been designed which con-
sist of CAE and Siamese Network. (ii) ROI images
are given as input to CAE capable of tuning image
features into compact network codes. In practice, the
proposed generalized hand vein transformation model
has been trained by explicitly reformulate the layers
as learning functions. We have trained a convolu-
tional deep auto-encoder with merged connections for
learning the TCM transformation and then trained an-
other similar auto-encoder for learning second trans-
formation IRT. In this way, we combine both models
and train an end-to-end CAE model from the orig-
inal image to final IRT transformed image. (iii) In
the last part, a Siamese network with triplet loss has
been trained and tested over the previously obtained
network transformed images. Thus, our deep learn-
ing vein recognition framework is highly generalized
Learning Domain Specific Features using Convolutional Autoencoder: A Vein Authentication Case Study using Siamese Triplet Loss
Network
779

for operating on either of palm vein or finger vein
databases. (iv) Deep networks usually require a lot
of training data, otherwise, they tend to over-fit. To
avoid that we pre-train the CNN in the Siamese net-
work using an autoencoder. (vi) Finally, the feature
level fusion of palm vein and finger vein modalities
has been performed to compare the performance.
2 PROPOSED VASCULAR
AUCTHENTICATION SYSTEM
At first, we extract the palm-vein and the four finger-
vein (index, middle, ring and little) region of inter-
est (ROI) seperatly from the given hand-vein images.
The ROI images are enhanced using contrast limited
adaptive histogram equalization (CLAHE). Following
this, an Autoencoder has been trained to learn Tex-
ture Code Matrix (TCM) and Image Ray Transform
(IRT) based encoding schemes, and finally, we use a
Siamese network trained using triplet loss for efficient
and accurate vein authentication.
2.1 Palm Vein and Finger Vein ROI
Extraction
We use the state-of-art algorithm (Bhilare et al., 2018)
for extracting the palm-vein ROI. Using this algo-
rithm we have been able to extract centre region of
hand as shown in figure 1 consistently.
We have modified (Bhilare et al., 2018) algorithm
in order to extract finger-vein ROI of index, middle,
ring and little fingers. Algorithm 1 and Algorithm 2
summarize the finger vein ROI extraction method.
2.2 Domain Specific Transformation
Learning using Autoencoders
In this work, the auto-encoder is inspired by U-Net
model which was used for segmentation task. We
have modified this model for learning image transfor-
mations, providing output similar to what it has been
trained on.
Network Architecture: The architecture of our deep
framework consists of two consecutive 3 × 3 convo-
lution layers with ReLU activation, each followed by
batch normalization. Thereafter a 2x2 max pooling
operation with stride 2 for down-sampling has been
used. At each down-sampling step, we double
the number of feature channels. Each step in up-
sampling path consists of 2×2 up-sampling operation
followed by a concatenation with the correspondingly
Algorithm 1: Finger-vein ROI extraction-Part 1.
Input: Acquired hand-vein image I, hand con-
tour C, finger tips T
I
, T
M
, T
R
and T
L
identified in algo-
rithm 1(palm-vein ROI extraction).
Output: Finger-vein ROIs corresponding to in-
dex f ROI
I
, middle f ROI
M
, ring f ROI
R
and little
f ROI
L
fingers
1: for i ∈ I, M, R,L do
2: T ← current finger f
i
3: T
Le ft
← adjacent finger on left side of T
4: T
Right
← adjacent finger on right side of T
5: [B
L
, B
R
] = FINDFINGERBASE-
POINTS(T, T
Le ft
, T
Right
,C) . Identify base points
for each finger f
i
6: Join B
L
and B
R
7: Rotate I such that B
L
B
R
is horizontal
8: Crop rectangular region with top T
i
and base
B
L
B
R
9: Perform morphological erosion to obtain
f ROI
i
10: end for
11: function FINDFINGERBASE-
POINTS(T, T
Le ft
, T
Right
,C)
12: if T
Le ft
6= φ and T
Right
6= φ then . Middle and
ring finger
13: P
Re f
= T
Le ft
. reference point
14: P
End
= T . end point
15: B
L
= FINDPEAK(P
Re f
, P
End
,C) . find the
point along the hand contour C between P
Re f
and
P
End
with maximum distance from P
Re f
16: P
Re f
= T
Right
17: P
End
= T
18: B
R
= FINDPEAK(P
Re f
, P
End
,C)
19: else if T
Le ft
= φ then . Little and ring finger
of left and right hand respectively
20: P
Re f
= T
Right
21: P
End
= T
22: B
R
= FINDPEAK(P
Re f
, P
End
,C)
23: B
L
= FINDEQUIDISTANTBASE-
POINT(T, B
R
,C)
24: else . Little and ring finger of right and left
hand respectively
25: P
Re f
= T
Le ft
26: P
End
= T
27: B
L
= FINDPEAK(P
Re f
, P
End
,C)
28: B
R
= FINDEQUIDISTANTBASE-
POINT(T, B
L
,C)
29: end if
30: return B
L
, B
R
31: end function
32: function FINDPEAK(P
Re f
, P
End
,C)
33:
p
max
= argmax
p
DISTANCE(P
Re f
,C
p
) (1)
ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods
780

Figure 2: Autoencoder (CAE) Architecture.
Algorithm 2: Finger-vein ROI extraction-Part 2.
Input: Acquired hand-vein image I, hand con-
tour C, finger tips T
I
, T
M
, T
R
and T
L
identified in algo-
rithm 1(palm-vein ROI extraction).
Output: finger-vein ROIs corresponding to index
f ROI
I
, middle f ROI
M
, ring f ROI
R
and little f ROI
L
fingers
34: return C(p
max
)
35: end function
36: function FINDEQUIDISTANTBASE-
POINT(T, B
1
,C)
37: x
1
= INDEXOF(C(B
1
)) . index of finger base
B
1
along the contour
38: x
2
= INDEXOF(C(T )) . index of finger tip T
along the contour
39: if x
1
> x
2
then
40: B
2
= C(x
2
− (x
1
− x
2
)) . B
1
and B
2
are
equidistant from T along the contour
41: else
42: B
2
= C(x
2
+ (x
2
− x
1
)
43: end if
44: return B
2
45: end function
feature map from the contracting path and reduction
in the number of output channels by the factor of 2,
and two 3 × 3 convolutions, each followed by batch
normalization and a ReLU. At the final layer, a 1 × 1
convolution is used to match the output image chan-
nels. In total the network consists of 15 convolution’s
layers and also avoids gradient problem. To allow
concatenation of layers it is important to select the in-
put tile size such that all 2×2 max-pooling operations
are applied to a layer with an even x- and y-size. Fig-
ure 2 shows the network architecture of image trans-
formation model.
Network Training: For training the autoencoders,
we first created ground truth by performing transfor-
mation operation namely TCM on 600 palm vein and
2400 finger vein samples and thereafter we applied
another transformation namely IRT for getting ray
tracing image features. We have used mean squared
error as loss function and RMSPROP optimizer with
default parameters. The detailed description of the
network training is as follows:
An autoencoder is trained to learn the TCM oper-
ation on the original image (palm vein or finger vein).
For that, the original image is shown to the network
and asked to generate the TCM of that image. Then a
second auto-encoder is trained which takes TCM im-
age as input and the network has to learn IRT opera-
tion on TCM images. This autoencoder is also trained
on the same number of images. Finally, we merged
these two autoencoder models (end to end) from orig-
inal to final transformation to create one deep autoen-
coder whose final task is to output the IRT image
from the original image itself as shown in Figure 3.
Thus, the combined autoencoder is then fine-tuned on
600 palm vein and 2400 finger vein images. The vi-
sual feature based appearance for palm vein and finger
vein is depicted in Figure 4, which clearly highlights
the curvilinear structures.
2.3 Siamese Matching Network
In order to match the multi-channel features, we train
a Siamese network using a triplet loss function. The
network consists of a Fused Feature Extractor (FFE)
which gives us the feature embedding for all 5 traits
of any single subject. Over these embedding, we ap-
ply triplet loss to train the FFE. Once the network is
trained, we match the samples using L
2
distance be-
tween the fused feature vectors obtained from FFE.
Triplet Loss Function: The feature embedding from
the FFE should be similar for a particular subject and
dissimilar for all different subjects. To ensure the pre-
vious assumption, we train our network using triplet
Learning Domain Specific Features using Convolutional Autoencoder: A Vein Authentication Case Study using Siamese Triplet Loss
Network
781

Figure 3: Combined Autoencoder Architecture.
Figure 4: CNN based Image Transformation: (a) Palm Vein Samples, (b)Finger Vein Samples.
Figure 5: Fused Feature Network.
loss. The loss function (D(a, p, n)), based on the dis-
tances between anchor (a), positive (p) and negative
(n) embedding have been used to train the network as
defined below (where M = 32 from hinge loss):
D(a, p, n) =
1
2
max(0, M + L
2
2
(a, p)) − L
2
2
(a, n) (2)
Fused Feature Extractor (FFE): For extracting the
multi-modal fused embedding, we use a Fused Fea-
ture Extractor as shown in Figure 5. It contains 5
CNN models for five traits (palm, index finger, middle
finger, ring finger and little finger). The architectures
of palm CNN differs from that of 4 finger CNN’s, as
the size of palm images are different from that of fin-
ger images. But since the 4 fingers, can also be treated
as different modalities, the weights are not shared be-
tween these networks. The architectures of the palm
and finger CNN are shown in Table 1.
Each CNN gives us a (7 ∗ 7 ∗ 512) dimensional
feature vector that is flattened for fusion. These flat-
tened vectors are then concatenated. A fully con-
nected layer of 128 neurons is applied over the con-
catenated vector, giving us a 128 dimensional 5 trait
fused feature vector. This has been done to achieve
a feature level fusion using multi-channel network.
This will ensure that each of the individual networks
optimizes not only for their corresponding traits but
for multiple traits and too simultaneously in one shot.
It has been observed that such a network can achieve
a better generalization ability.
Network Training: Deep networks require a lot of
training data to learn the features. Since that luxury
we do not have in biometrics, to ensure that the fea-
tures learned are robust enough we first pre-train the
CNNs’ in the FFE using an autoencoder. This is an
efficient procedure to train a network when not much
data is available. Such an autoencoder can let network
to initially understand how to regenerate the same in-
put image by learning the holistic image generative
features. Later we remove the decoder part and train
the encoder to fulfil our objective to finally tune pa-
ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods
782

Figure 6: CNN Architecture for Palm Vein. Trained a 3-
channel autoencoder to pre-train the encoder of Siamese
network.
Figure 7: CNN Architecture for Finger Vein. Trained a
3-channel autoencoder to pre-train the encoder of Siamese
network.
rameters to learn the best discriminative features.
We have designed an autoencoder whose encoder
part is same as the CNN to be used in FFE. The in-
put to encoder being the multi-channel feature image
and output is a feature map of size (7 ∗ 7 ∗ 512). The
decoder has been fixed as a mirror of the CNN en-
coder, which will take the feature map from the en-
coder as an input and output the multi-channel fea-
ture image. We have trained this autoencoder over
the gallery dataset ensuring us that the encoder has
learned all the generative features present in the im-
ages. Once the autoencoder is trained, we discard the
decoder and save the weights of the encoder and use
it as a pre-trained weights in the FFE as shown in Fig-
ure 6 and Figure 7. Later we fine tune the FFE for
the discrimination task using the triplet loss (as dis-
cussed above). Once the Siamese model is trained,
we combine both Autoencoder and Siamese network
and train the whole network in an end-to-end fashion.
Table 1: CNN Architecture for Palm Vein.
Palm Vein
Input Size 144 × 144 × 3
CONV 32(9,9), Stride (1,1), Activation: Relu
MaxPool(2,2), Stride (2,2)
CONV 64(7,7), Activation: Relu, Padding(1,1)
MaxPool(2,2), Stride (2,2)
CONV 128(5,5), Activation: Relu, Padding(1,1)
MaxPool(2,2), Stride (2,2)
CONV 128(3,3), Activation: Relu, Padding(1,1)
MaxPool(2,2), Stride (2,2)
CONV 256(3,3), Activation: Relu, Padding(1,1)
CONV 512(3,3), Activation: Sigmoid, Padding(0,0)
Output Size 7 × 7 × 512
Finger Vein
Input Finger Vein Size 144 × 28 × 3
CONV 32(9,9), Activation: Relu, Padding(1,1)
MaxPool(2,1), Stride (2,1)
CONV 64(7,7), Activation: Relu, Padding(1,1)
MaxPool(2,1), Stride (2,1)
CONV 128(5,5), Activation: Relu, Padding(1,1)
MaxPool(2,2), Stride (2,2)
CONV 128(3,3), Activation: Relu, Padding(1,1)
MaxPool(2,2), Stride (2,2)
CONV 256(3,3), Activation: Relu, Padding(1,1)
CONV 512(3,3), Activation: Sigmoid, Padding(0,1)
Output Finger Vein Size 7 × 7 × 512
3 EXPERIMENTAL ANALYSIS
In distinguishing experiments, the performance of the
proposed method has been evaluated in terms of EER
(Equal Error Rate), and DI (Decidability Index).
Database Specifications: The proposed system has
been tested on publicly available Multispectral CA-
SIA palm print database (43, ), which consist of 6 im-
ages per subject for all the palm-vein and finger-vein
traits. The left and right-hand samples from a sub-
ject are considered belonging to separate individuals,
which resulted in 200 subjects. For each subject, the
first three samples are considered as the gallery and
the remaining as the probe.
Experiment 01: In the first experiment, the indi-
vidual performance of finger vein and palm vein
modalities have been computed. The corresponding
ROC characteristics for palm vein and finger vein
are shown in Figure 8. The matching between the
gallery and probe samples are computed resulting in
1800 genuine and 358200 imposter scores. The val-
ues of EER and DI corresponding to the best perfor-
mance for each modality have been highlighted in Ta-
ble 2. To make a fair comparison, six state-of-the-art
methods are reported. From table 2, we can make
the following inferences, Firstly, palm vein samples
achieved superior performance than any of the fin-
Learning Domain Specific Features using Convolutional Autoencoder: A Vein Authentication Case Study using Siamese Triplet Loss
Network
783

Table 2: Comparative Analysis (P: Palm; F: FingerVein; (+) : Fusion; 4F : All four fingervein).
Evaluation Parameters
Dataset CASIA
Approach P/F P+1F 4F P+4F DI
WLD (Huang et al., 2010) 6.08 6.11 - 5.23 -
MPC (Choi et al., 2009) 4.54 4.92 - 3.32 -
RLT (Miura et al., 2004) 17.17 18.56 9.00 13.98 -
NMRT (Zhou and Kumar, 2011) 9.58 7.88 - 6.06 -
Hessian phase (Zhou and Kumar, 2011) 9.39 9.09 - 7.24 -
Proposed Palm vein 4.12 4.02 - - 2.50
Proposed Finger vein 8.50 - 6.45 - 2.15
Proposed Fused All - - - 3.33 3.05
ger vein traits because of bigger and better ROI. Sec-
ondly, the proposed deep network clearly outperforms
the state-of-the-art for palm-vein. The EER values
of 4.125 % and 4.02% have been obtained over CA-
SIA database. Among the individual performance of
finger-vein traits, the best EER value of 8.50 % has
been achieved by -middle finger. Therefore it is clear
that the palm-vein is more discriminatory trait as com-
pared to the finger-vein traits.
Experiment 02: The second experiment studies
about the feature level fusion of palm vein and finger
vein modalities. In addition to this, the combined per-
formance of 4 fingers has also been evaluated. A low
EER value of 3.33% has been achieved with feature
level fusion of palm vein and four finger veins, which
is superior to the EER values obtained from any of the
5 state of art methods as well as the proposed palm
vein or finger vein individual networks. This justifies
the strength of network learned fused feature repre-
sentation. The detail description of other parameters
is given in Table 2. The corresponding ROC charac-
teristics and genuine Vs imposter score distribution
graph for fused (palm vein, 4 finger vein) and fused
finger vein are shown in Figure 9 and Figure 10.
4 CONCLUSIONS
In this work, we have proposed a novel end-to-end
deep network design by combining domain specific
knowledge and deep learning representation. The var-
ious challenging issues related to vein biometrics have
been addressed suitably. The fixed size ROI images
have been given to an end-to-end CAE augmented
with Siamese network trained using triplet loss for
vein recognition. Finally, in order to utilize the in-
formation present in the whole Palmer region of the
hand, feature level fusion of palm vein and finger vein
has been performed. As a part of future work, we
will try to incorporate CNN based ROI segmentation
Figure 8: ROC based Performance Analysis for Palm Vein
and Finger Vein.
Figure 9: ROC based Performance Analysis for Fused Fea-
tures.
and ROI enhancement network to further improve the
recognition performance of the proposed system.
ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods
784

Figure 10: Genuine Vs Imposter Score Distribution (All
traits fused together).
REFERENCES
Casia m.: Palmprint v1 database [online]. available:.
http://www.cbsr.ia.ac.cn/ms palmprint.
Ahonen, T., Hadid, A., and Pietikainen, M. (2006). Face
description with local binary patterns: Application to
face recognition. IEEE transactions on pattern analy-
sis and machine intelligence, 28(12):2037–2041.
Bhilare, S., Jaswal, G., Kanhangad, V., and Nigam, A.
(2018). Single-sensor hand-vein multimodal bio-
metric recognition using multiscale deep pyrami-
dal approach. Machine Vision and Applications,
29(8):1269–1286.
Choi, J. H., Song, W., Kim, T., Lee, S.-R., and Kim, H. C.
(2009). Finger vein extraction using gradient normal-
ization and principal curvature. In Image Processing:
Machine Vision Applications II, volume 7251, page
725111. International Society for Optics and Photon-
ics.
Cummings, A. H., Nixon, M. S., and Carter, J. N. (2011).
The image ray transform for structural feature detec-
tion. Pattern Recognition Letters, 32(15):2053–2060.
Fang, Y., Wu, Q., and Kang, W. (2018). A novel finger vein
verification system based on two-stream convolutional
network learning. Neurocomputing.
Huang, B., Dai, Y., Li, R., Tang, D., and Li, W. (2010).
Finger-vein authentication based on wide line detec-
tor and pattern normalization. In Pattern Recognition
(ICPR), 2010 20th International Conference on, pages
1269–1272. IEEE.
Jaswal, G., Nigam, A., and Nath, R. (2017). Deepknuckle:
revealing the human identity. Multimedia Tools and
Applications, 76(18):18955–18984.
Kimura, T., Makihara, Y., Muramatsu, D., and Yagi,
Y. (2015). Single sensor-based multi-quality multi-
modal biometric score database and its performance
evaluation. In Biometrics (ICB), 2015 International
Conference on, pages 519–526. IEEE.
Miura, N., Nagasaka, A., and Miyatake, T. (2004). Feature
extraction of finger-vein patterns based on repeated
line tracking and its application to personal identifi-
cation. Machine vision and applications, 15(4):194–
203.
Nigam, A., Tiwari, K., and Gupta, P. (2016). Multiple
texture information fusion for finger-knuckle-print au-
thentication system. Neurocomputing, 188:190–205.
Patil, I., Bhilare, S., and Kanhangad, V. (2016). Assess-
ing vulnerability of dorsal hand-vein verification sys-
tem to spoofing attacks using smartphone camera. In
Identity, Security and Behavior Analysis (ISBA), 2016
IEEE International Conference on, pages 1–6. IEEE.
Qin, H. and El-Yacoubi, M. A. (2017). Deep representation-
based feature extraction and recovering for finger-
vein verification. IEEE Transactions on Information
Forensics and Security, 12(8):1816–1829.
Ramalho, M., Correia, P. L., Soares, L. D., et al. (2011).
Biometric identification through palm and dorsal hand
vein patterns. In EUROCON-International Confer-
ence on Computer as a Tool (EUROCON), 2011
IEEE, pages 1–4. IEEE.
Taigman, Y., Yang, M., Ranzato, M., and Wolf, L. (2014).
Deepface: Closing the gap to human-level perfor-
mance in face verification. In Proceedings of the IEEE
conference on computer vision and pattern recogni-
tion, pages 1701–1708.
Umer, S., Dhara, B. C., and Chanda, B. (2016). Tex-
ture code matrix-based multi-instance iris recognition.
Pattern Analysis and Applications, 19(1):283–295.
Veluchamy, S. and Karlmarx, L. (2016). System for multi-
modal biometric recognition based on finger knuckle
and finger vein using feature-level fusion and k-
support vector machine classifier. IET Biometrics,
6(3):232–242.
Wang, J.-G., Yau, W.-Y., Suwandy, A., and Sung, E. (2008).
Person recognition by fusing palmprint and palm vein
images based on “laplacianpalm” representation. Pat-
tern Recognition, 41(5):1514–1527.
Xie, C. and Kumar, A. (2017). Finger vein identification us-
ing convolutional neural network and supervised dis-
crete hashing. In Deep Learning for Biometrics, pages
109–132. Springer.
Yang, J. and Zhang, X. (2012). Feature-level fusion of
fingerprint and finger-vein for personal identification.
Pattern Recognition Letters, 33(5):623–628.
Yang, W., Huang, X., Zhou, F., and Liao, Q. (2014). Com-
parative competitive coding for personal identification
by using finger vein and finger dorsal texture fusion.
Information sciences, 268:20–32.
Zhou, Y. and Kumar, A. (2010). Contactless palm vein iden-
tification using multiple representations. In Biomet-
rics: Theory Applications and Systems (BTAS), 2010
Fourth IEEE International Conference on, pages 1–6.
IEEE.
Zhou, Y. and Kumar, A. (2011). Human identification using
palm-vein images. IEEE transactions on information
forensics and security, 6(4):1259–1274.
Learning Domain Specific Features using Convolutional Autoencoder: A Vein Authentication Case Study using Siamese Triplet Loss
Network
785