Pattern Recognition 120 (2021) 108122

Contents lists available at ScienceDirect

Pattern Recognition
journal homepage: www.elsevier.com/locate/patcog

A survey on dorsal hand vein biometrics

Wei Jia a,b,c, Wei Xia b, Bob Zhang d,∗, Yang Zhao a, Lunke Fei e, Wenxiong Kang f, Di Huang g,
Guodong Guo h
a
Key Laboratory of Knowledge Engineering with Big Data, Ministry of Education, Hefei University of Technology, Hefei 230009, China
b
School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China
c
Anhui Province Key Laboratory of Industry Safety and Emergency Technology, Hefei 230009, China
d
PAMI Research Group, Dept. of Computer and Information Science, University of Macau, Macau 999078, China
e
School of Computers, Guangdong University of Technology, Guangzhou, 510006, China
f
School of Automation Science and Engineering, South China University of Technology, Guangzhou, 510006, China
g
School of Computer Science and Engineering, Beihang University, Beijing, 100083, China
h
Department of Computer Science and Electrical Engineering, West Virginia University, West Virginia, 26506, United States

a r t i c l e i n f o a b s t r a c t

Article history: Biometrics technology is one of the most important and effective solutions for personal authentication.
Received 20 August 2019 In recent years, as one of the emerging biometrics technologies, dorsal hand vein (DHV) biometrics has
Revised 18 May 2021
received a lot of attention. In fact, DHV biometrics has been studied for more than 30 years, during which
Accepted 20 June 2021
different problems related to DHV recognition have been addressed. In this paper, we conduct a compre-
Available online 28 June 2021
hensive survey on the state-of-the-art in DHV biometrics. Nearly all important aspects of DHV biomet-
Keywords: rics have been summarized, including the developmental history, data acquisition, databases, preprocess-
Biometrics ing algorithms, feature extraction and matching algorithms, information fusion schemes and commercial
Dorsal hand vein products. We also discuss the challenges and future directions in DHV biometrics research.
Survey
© 2021 Elsevier Ltd. All rights reserved.

1. Introduction High Recognition Performance [2]: The features within DHV of

each individual are unique and distinctive. Furthermore, the DHV
Biometrics technology-based person authentication is becoming pattern is robust to environmental changes, pollution, scarring
more and more important in our digital society. Many human traits and physical breakage. Thus, DHV-based biometrics systems can
have been successfully used for the biometrics, including face, iris, achieve a high recognition accuracy.
fingerprint, voice and signature, etc. Besides these, researchers in High Ability of Liveness Detection [3]: DHVs are sensed by in-
both industry and academia are still attempting to seek other hu- frared imaging, and they can only be captured on a live body.
man traits for effective person authentication. Therefore, the DHV biometrics has a natural ability for liveness de-
In recent years, dorsal hand vein (DHV) recognition, as an tection.
emerging biometrics technology, has received much attention. It is High Anti-Counterfeit Performance [2 4]: As DHVs are inside the
well known that DHV recognition belongs to the family of hand- body, it is difficult to forge them, thereby leading to a high anti-
based biometrics. As shown in Fig. 1, from the front and back view counterfeit performance.
of a hand, a lot of physiological properties can be extracted for High Acceptability [2]: The acquisition of the DHV is simple and
the purpose of biometrics, such as fingerprint, hand geometry, in- fast, and the users do not need to worry about the health issues.
ner knuckle print, palmprint, palm vein, finger vein, fingernail bed, In the past 30 years, great progress has been made in the re-
finger knuckle print, and DHV [1]. In Fig. 2, we can see that DHV search on DHV biometrics. In this paper, we conduct a compre-
recognition is the technology using automatic analysis of the sub- hensive survey on the state-of-the-art in DHV biometrics. Specif-
cutaneous venous network on the back of the hand for the appli- ically, we present an overview of DHV biometrics, including the
cations of person authentication. As often mentioned in some lit- development history, data acquisition, databases, preprocessing al-
erature [2–4], DHV biometrics is attractive due to its following dis- gorithms, feature extraction and matching algorithms, information
tinguished advantages. fusion schemes, and commercial products. We also discuss the
challenges and future directions in DHV biometrics research.
∗
We organize the rest of this paper as follows. Section 2 presents
Corresponding author.
the historical developments of DHV biometrics. Section 3 sum-
E-mail address: [email protected] (B. Zhang).

https://doi.org/10.1016/j.patcog.2021.108122
0031-3203/© 2021 Elsevier Ltd. All rights reserved.
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 1
Infrared sub-division [16].

Division name Abbreviation Wavelength (μm)

Near-infrared NIR 0.7–1.4

Short-wavelength infrared SWIR 1.4–3
Mid-wavelength infrared MWIR 3–8
Long-wavelength infrared LWIR 8–15
Far-infrared FIR 15–1,000
Fig. 1. Front and back view of a hand [1].

tion feature of DHV. In 2015, Huang et al. [11] presented a DHV

recognition approach through matching local features of multiple
sources, which achieved a high recognition accuracy. In 2016, it is
the first time that Li et al. [12] introduced deep learning technique
for DHV recognition. In their work, the recognition accuracies of
several popular Convolutional Neural Networks (CNN) were com-
pared. In 2017, Wang et al. [2] designed a transmission mode-based
DHV acquisition device and established a DHV image database, and
then proposed a method for both identity recognition and gender
recognition using fine-tuned Deep Neural Network (DNN) models.
In 2018, Gupta et al. [13] presented a multimodal biometric sys-
tem by fusing the information of slap fingerprints, DHV and hand
geometry for robust person authentication. In 2019, Zhong et al.
[14] proposed a multimodal biometric system, in which the fusion
of DHV and palmprint has been studied in depth. In 2019, Wang
et al. [15] introduced the techniques of Generative Adversarial Net-
works (GAN) and model compression to address some problems of
deep learning-based DHV recognition.

3. Image acquisition, databases and preprocessing

Fig. 2. The dorsal venous network of the hand. 3.1. DHV image acquisition

According to different criteria, DHV image acquisition modes

marizes the image acquisition, databases, and several preprocess- can be divided into different categories.
ing algorithms including the Region of Interest (ROI) extraction al-
gorithms, as well as image quality assessment and enhancement 3.1.1. Acquisition modes according to electromagnetic spectrum
algorithms. Section 4 provides a new taxonomy for DHV recog- DHV images are usually captured by infrared imaging. In [16],
nition methods and summarizes existing recognition methods. Gade et al. depicted the electromagnetic spectrum with sub-
Section 5 summarizes information fusion schemes. Section 6 intro- divided infrared spectrum. As illustrated in Fig. 4, infrared ra-
duces the commercial products. Section 7 highlights the challenges diation lies between visible light and microwaves within the
and future directions in DHV biometrics research. Finally, we con- wavelength spectrum of 0.7–1,0 0 0 μm. The infrared spectrum
clude the paper in Section 8. can be divided into several spectral regions. There exist different
sub-division schemes in different scientific fields, but a common
2. Historical developments of DHV biometrics scheme is shown in Table 1 [16], which includes near-infrared
(NIR), short-wavelength infrared (SWIR), mid-wavelength infrared
Fig. 3 depicts the development history of DHV biometrics, in (MWIR), long-wavelength infrared (LWIR) and far-infrared (FIR). As
which several major milestones are presented. In 1985, Joseph Rice, shown in Fig. 5, according to the sub-divided infrared spectrum,
an automation controls engineer at Kodak’s Annesley Factory, in- the DHV image acquisition modes can be divided into FIR-based,
vented hand vein pattern recognition, and he filed a patent, which NIR-based, and multispectral-based, respectively.
exploited hand vein for person authentication in the United States The principle of the FIR imaging-based methods is as follows
[5]. In 1991, MacGregor et al. [6] presented a DHV recognition [17,152]. A human body emits infrared radiation, thus, by using
system for personnel authentication, which was named as ‘Vein a thermal camera with detector sensitivity, an image showing the
Check’. Later, a joint venture created by Cambridge Consultants and heat distribution of the human body can be generated. Human su-
British Technology group commercialized the ‘Vein Check’ system. perficial veins have higher temperature than the surrounding tis-
In 1995, Cross and Smith acquired the DHV by the thermographic sues. Therefore, via thermal imaging, the images containing the
imaging technology for biometric applications [7]. In 1997, Alex heat distribution of body parts can clearly display the structure of
Hwansoo Choi, the co-founder of BK Systems, launched the first the desired vein patterns, as shown in Fig. 5(a). In addition, be-
commercial DHV recognition products: BK-10 0 [8]. In 20 0 0, some cause FIR imaging forms images using the infrared radiation emit-
members of BK Systems founded the Techsphere Co. Ltd, and re- ted from the objects, no external lighting is required.
leased a new commercial product of DHV recognition: VP-II [8]. The principle of the NIR imaging-based methods is as fol-
After 20 0 0, there are several representative works on DHV lows [17,152]. The haemoglobin is the pigment in the blood
recognition. In 2004, Lin et al. [9] presented a DHV recognition ap- which is primarily composed of iron, which carries the oxygen.
proach using the thermal images. In 2009, Kumar et al. [10] pro- Haemoglobin is known to absorb NIR light. The blood in veins ob-
posed a multi-feature-based approach for person authentication viously contains a higher amount of deoxygenated haemoglobin as
by combining the knuckle shape information and the triangula- compared to blood in arteries. Deoxygenated hemoglobin is the

2
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Fig. 3. Major milestones in the development history of DHV biometrics.

Fig. 7. The acquisition modes in NIR imaging. (a) Reflection mode, (b) Transmission
Fig. 4. The electromagnetic spectrum with sub-divided infrared spectrum [16].
mode.

be rendered visible with the help of an NIR light source in combi-

nation with an NIR-sensitive image sensor. As shown in Fig. 5(b),
under NIR illumination and after imaging, vessels appear as dark
structures, meanwhile, the surrounding tissue appears bright due
to a much lower light absorption coefficient in NIR spectrum.
Wang et al. [17] investigated FIR and NIR imaging-based palm
vein and DHV acquisitions for the purposes of biometrics. Their ex-
Fig. 5. The acquisition modes according to electromagnetic spectrum. (a) DHV im- perimental results showed that FIR imaging is sensitive to human
age obtained by FIR-based acquisition [9], (b) DHV image obtained by NIR-based body condition and ambient conditions while NIR imaging has bet-
acquisition [17], (c) DHV images obtained by multispectral-based acquisition [23].
ter tolerance to environment and body condition changes. For FIR-
based DHV recognition, a lot of studies have been conducted [9,18–
21]. However, because the cost of FIR capture devices is usually
high and the image quality of FIR imaging is not stable, most re-
searchers focused on NIR-based DHV recognition in recent years.
Only a few works have been conducted for multispectral DHV
recognition [22,23], and the main aim of these works is to inves-
tigate the best band of spectra for DHV recognition. The authors
of [22,23] designed a multispectral DHV image acquisition device,
and collected a multispectral image database. Their experimental
results demonstrated that the DHV recognition system can work
well on the infrared spectra with the wavelength from 520nm to
1040nm, and the spectrum of 880nm or 890nm is the best choice
for recognition

3.1.2. Acquisition modes for NIR imaging

Usually, NIR imaging based DHV acquisition has two modes, i.e.,
transmission mode and reflection mode [24]. In these two modes,
the acquisition device first uses an LED array to emit the NIR light
to the hand, and then uses a CCD camera sensitive to NIR light
to capture the DHV image. In order to block all undesired visible
light emitted by external sources, a NIR filter should be attached in
front of the CCD camera. For the reflection mode, the CCD camera
Fig. 6. Absorption spectra of hemoglobin in NIR spectrum. and the illumination LED array are in the same side as shown in
Fig. 7(a). For the transmission mode [2,15,25–27], the LED array is
under the palm of the hand, and the CCD camera is placed above
form of hemoglobin without the bound oxygen. The absorption the back of hand to capture the light that passes through the hand
spectra of oxyhemoglobin (Oxy Hb) and deoxyhemoglobin (Deoxy as shown in Fig. 7(b). Fig. 8 shows several DHV images captured by
Hb) differ. As shown in Fig. 6, Oxy Hb and Deoxy Hb absorb NIR transmission mode and reflection mode, respectively. Kauba et al.
light equally at 805 nm. In wavelength less than 805 nm, the Oxy [24] made a comparison between transmission and reflection ac-
Hb has lower absorption than Deoxy Hb, while above 805 nm has a quisition modes. Compared to reflection mode, they found that
stronger absorption. Thus, the vascular pattern inside the hand can transmission mode is more robust against the influence of ambi-

3
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

listed in Table 2. It can be seen that most databases were cap-

tured by using NIR imaging and the reflection mode, which indi-
cates that NIR imaging with reflection mode is the most commonly
used imaging method. In real applications, taking public health
into consideration, it is a better choice to use touchless mode to
collect biometrics data. However, only 7 databases were collected
by the touchless mode. From Table 3, we can see that the number
of subjects in most of the databases is less than 300, which means
the scales of most DHV databases are small. In order to simulate
real application scenarios, the data should be captured across dif-
ferent sessions. However, only 12 databases were captured in two
sessions including 2 public databases.

3.3. Region of interest (ROI) extraction algorithms

ROI extraction is a key step in DHV recognition. As shown

Fig. 8. DHV images captured by reflection mode (a) [24] and transmission mode in Figs. 11 and 12, to extract ROI from the open hand DHV im-
(b) [24].
ages, researchers usually used the information of valley points be-
tween the fingers to align open hand images [9,17,18,27]. Yakno
et al. [32] presented a discussion on ROI extraction algorithms and
proposed an improved algorithm for larger ROI extraction. Damak
et al. [59] investigated a contour profile distance distribution-based
method for ROI extraction. As shown in Figs. 13 and 14, to extract
ROI from clenched hand image, researchers usually used the infor-
mation of valley points [23,25,40,60,61] or tip points [10] between
the knuckle to align clenched hand images. From Figs. 11–14, it can
Fig. 9. Touch mode and touchless mode. (a) Touch mode, (b) Touchless mode.
be seen that a general ROI extraction algorithm usually includes
the following main steps: (1) converting the gray hand image into
a binary image by a segmentation algorithm; (2) performing hand
boundary tracking; (3) calculating the distance profile between the
contour points (the points located on the contour of the hand) and
the reference point. Generally, the mid-point of the wrist is treated
as the reference point; (4) locating the ROI based on the detected
points; (5) cropping the ROI sub-image.

Fig. 10. Acquisition modes according to hand pose. (a) Open hand mode, (b) 3.4. Image quality assessment and enhancement algorithms
clenched hand mode.
For image-based biometric systems, poor image quality often
leads to bad recognition performance. In infrared imaging, image
ent light and the hand surface conditions, but needs a higher light
acquisition is often affected by luminosity and thickness of the
intensity, thus results in consuming more power.
back of the hand skin, which leads to bad image quality such as
poor contrast, non-uniform gray level and noise. Thus, image qual-
3.1.3. Acquisition modes according to touch or touchless
ity assessment and image enhancement for DHV recognition have
As shown in Fig. 9, according to whether the hand touches
been investigated.
the capture device or not, the acquisition modes can be di-
Several DHV image quality assessment algorithms have been
vided into two categories: touch-based acquisition [23,25,28,29]
proposed. Wang et al. [39] proposed to use the method of total
and touchless-based acquisition [10,30,31].
variation (TV) to evaluate the DHV image noise. Zhao et al. [62] ex-
ploited “mean gray value” and “gray variance” as the image quality
3.1.4. Acquisition modes according to hand pose index. Hartung et al. [63] proposed a DHV image quality estimation
As shown in Fig. 10, according to the hand pose, the acquisition algorithm using global and local features based on Grey Level Co-
modes can be classified into two categories, i.e., open hand mode Occurrence Matrix (GLCM) and optionally available metadata. Xu
[18,26,32], and clenched hand mode [10,11,22,23,30]. et al. [64] proposed a DHV image quality assessment method based
on two-dimensional entropy. Cui et al. [65] performed DHV image
3.2. DHV image databases quality assessment using four parameters including gray variance,
information entropy, cross point and area of effect, and then fusing
Databases are the foundation for biometrics research. For the them according to weights. Wang et al. [66] selected five parame-
research of DHV recognition, one of the main problems is the ters as the reference standard for DHV image quality assessment
lack of public DHV image databases, therefore, researchers usually including effective region size, contrast, definition, geometric devi-
constructed databases by themselves. Table 2 summarizes the de- ation and rotation angle. Gupta et al. [57] used two well-known al-
tails of DHV image databases, in which a total of 37 DHV image gorithms to estimate the vein quality including gradient based vein
databases is listed. It should be noted that because most databases quality estimation and local range of an image. Wang et al. [2] pro-
have no names, we name them using the abbreviation of the first posed a sharpness-based vein quality estimation method, in which
author’s affiliation. the sharpness score of vein images is calculated with the improved
From Table 2, it can be seen that only 5 databases are public in- Fast Image SHarpness (FISH) by adding the contrast information
cluding NCUT [38], GPDS [37], Bosphorus [45], PROTECTVein [24], (CFISH) with appropriate weight set. Kauba et al. [24] used 3 im-
and JLU [162]. Table 3 provides other statistics of the databases age contrast metrics for contrast and quality assessment including

4
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 2
The list of DHV image databases.

Reflection
(R) or Number Image
Trans- Number of number Total
mission of indi- Number capture Time of each image
References Name Spectrum (T) Touch viduals of hands session interval session number Year

Lin et al. [9] NCU FIR N/A N/A 32 32 3 1 week 10 960 2004
Badawi et al. [33] TU NIR R Yes 500 1000 1 N/A 5 5000 2006
Wang et al. [17] NTU FIR N/A Yes 30 60 2 N/A 9 540 2006
NIR R Yes 150 300 1 N/A 3 900
Wang et al. [34] BIT NIR R Yes 100 100 1 N/A 5 500 2008
Kumar et al. [10] IITD NIR R No 100 100 1 N/A 3 300 2009
Wang et al. [35,36] TJU NIR R Yes 108 108 1 N/A 7 756 2009
Ferrer et al. [37] GPDS (Public) NIR R Yes 102 102 2 1 week 5 1020 2009
Wang et al. [38,39] NCUT (Public) NIR R Yes 102 204 1 N/A 10 2040 2009
Yuan et al. [40] BJU NIR R Yes 30 30 1 N/A 20 600 2010
Lu et al. [41] HU NIR R Yes 100 100 1 N/A 10 1000 2010
Hsu et al. [42] NDU NIR R No 214 214 2 1 month 10 4280 2011
Chen et al. [23] HKPU Multispectral R Yes 211 422 1 N/A 5 × 53 11130 2011
Ramalho et al. [43] IUL NIR R Yes 70 140 1 N/A 10 1400 2011
Bu et al. [44] HIT NIR R No N/A 180 1 N/A 20 × 3 10800 2011
Yuksel et al. [45] Bosphorus NIR R Yes 100 100 4 N/A 3 1575 2011
(Public)
Kang et al. [46] SCUT NIR R Yes 60 120 3 7 days 1 360 2012
Meng et al. [47] FU NIR R Yes 168 168 1 N/A 3 504 2013
Park et al. [48] KU NIR R Yes 100 100 1 N/A 13 1300 2013
Lee et al. [49] CNA NIR R No 308 308 2 1 month 10 6160 2014
Faundez-Zanuy et al. Tecnocampus NIR; FIR R; N/A Yes 100 100 5 4 weeks 2 1000; 2014
[50] 1000
Neves et al. [26] IST FIR N/A Yes 30 60 1 N/A 5 300 2014
Castro-Ortega et al. UPT NIR R Yes 38 76 1 N/A 12 912 2015
[51]
Gupta et al. [52] IITK-1 NIR R Yes 1030 1030 1 N/A 4 4120 2015
Gruschina et al. [53] VeinPLUS NIR R; T Yes 107 214 1 N/A 3 1213 2015
Zhu et al. [25] NUAA NIR T Yes 43 43 1 N/A 9 387 2015
Cao et al. [28] MyPalmVein1.0 NIR T Yes 270 270 1 N/A 3 810 2015
MyPalmVein2.0 NIR T Yes 25 25 1 N/A 5 125 2015
Raghavendra et al. [54] CUC NIR R Yes 50 100 2 N/A 5 1000 2015
Lin et al. [55] HHUT FIR N/A Yes 100 100 1 N/A 30 3000 2015
Khan et al. [56] UM NIR R Yes 300 300 1 N/A 10 3000 2015
Joardar et al. [30] JU-NIR-V1 NIR R No 30 30 1 N/A 10 300 2016
Joardar et al. [31] JU-NIR-V2 NIR R No 60 60 1 N/A 10 600 2016
Gupta et al. [57] IITK-PDV NIR R No 150 300 2 8 days 2-3 1500 2016
Wang et al. [2,58] CUMT NIR T Yes 200 200 2 10 days 5 2000 2017
Gupta et al. [13] IITK-2 NIR R Yes 200 400 2 20 days 2-3 2000 2018
Kauba et al. [24] PROTECTVein NIR R; T Yes 40 80 1 N/A 30 2400 2018
(Public)
Zhong et al. [14] XJUT NIR R Yes 57 113 1 N/A 10 1130 2019
Liu et al. [162] JLU (Public) NIR R Yes 736 736 1 N/A 5 3680 2020

Table 3
Some statistics of 37 DHV databases listed in Table 2.

Capture
Subjects sessions
NIR Reflection Touchless number ≥ number ≥
imaging mode mode 300 2

Number 31 31 7 5 12
Ratio 83.78% 83.78% 18.91% 13.51% 32.43%

Fig. 11. Open hand ROI cropping algorithm proposed by Lin et al. [9] and Wang et al. [17,18]. (a) A DHV image, (b) the boundary of the hand, (c) the distance profile between
the reference point and the contour points, (d) locating the ROI based on the detected points.

5
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Fig. 12. Open hand ROI cropping algorithm proposed by Wang et al. [27]. (a) DHV image, (b) binary image, (c) the boundary of the hand, (d) the angular point positioning
image, (e) the ROI region.

Fig. 13. Clenched hand ROI cropping algorithm proposed by Meng et al. [60]. (a) Binary image, (b) the boundary of the hand, (c) distance profile between the reference
point and the counter points, (d) locating the ROI, (e) the extracted ROI image.

Fig. 14. Clenched hand ROI cropping algorithm proposed by Chen et al. [23]. (a) The original DHV image, (b) binary image, (c) the distance vector calculation, (d) the
invariant feature searching, (e) the ROI region.

Global Contrast Factor (GCF), Grey Level Entropy Score (GLES), and ent contrast enhancement algorithms on the performance of Scale-
Entropy Score (ES). In general, the studies for DHV image quality Invariant Feature Transform (SIFT) based DHV recognition system.
assessment are relatively inadequate. Furthermore, there lacks an They found that the use of contrast enhancement algorithms has
extensive performance evaluation of existing DHV image quality a positive effect on increasing the number of keypoints, but has
assessment algorithms. Thus, in the future, more efforts should be an obvious negative effect on keypoints matching which leads to
invested in improving the research on DHV image quality assess- worse recognition results. In order to eliminate the negative effect
ment. caused by contrast enhancement algorithm, Li et al. [89] proposed
In the DHV recognition filed, there are many works for im- a hierarchical matching strategy to thoroughly remove unreliable
age enhancement. Generally, these image enhancement algorithms matching. Wang et al. [2] proposed a “contrast enhancement-
have two main steps. The first one is to remove noise, and the sec- free” DHV recognition system, in which different feature extraction
ond one is to increase the contrast. For noise removal, the com- strategies were respectively applied to high quality and low qual-
monly used tools include median filter [18,7,54,67–70], Gaussian ity DHV images. For the high quality DHV images, no contrast en-
smoothing filter [71–73], Wiener filter [74,75], and wavelet trans- hancement algorithm was used. For the low-quality DHV images,
form [76,77], etc. In order to increase the contrast, a lot of algo- special strategy with contrast enhancement was designed to im-
rithms or tools were used including Guided filter (GF) [57,81,94], prove the discriminability of the feature and dismiss the negative
Bilateral filter (BF) [81], Intensity normalization (IN) [2], Histogram effect of contrast enhancement. The work of [2] indicates that it
Equalization (HE) [51,69,70,73,77–82], Adaptive Histogram Equal- may be not necessary to apply contrast enhancement to high qual-
ization (AHE) [10,81,89,90], DoG filter and HE (DHE) [2,89,90], Con- ity DHV images, and for low quality DHV image processing, the
trast limited histogram equalization (CLHE) [2], Contrast Limited contrast enhancement step is necessary. In our opinion, it is hard
Adaptive Histogram Equalization (CLAHE) [30,31,51,78,80,81,83– to say which image enhancement algorithm is the best. For differ-
86], Contrast Stretching (CS) [78,79], Gamma correction (GC) [2], ent kinds of DHV databases and different feature extraction meth-
Retinex and Adaptive Smoothing Filter (RASF) [87,88], Gray Level ods, different image enhancement algorithms may be needed. That
Slicing (GLS) [2,79,86], Homomorphic filter (HF) [2,86], High Fre- is to say, image enhancement for DHV image processing is still an
quency Filtering and HE (HHE) [85], Image Negative Enhancement unsolved problem. In the future, more extensive performance eval-
(INE) [2,79,86], Laplacian Sharpening (LS) [79,86], Phase congru- uations of existing DHV image enhancement algorithms should be
ency (PC) [151], Unsharp Masking (UM) [79,86], High Boost Fil- conducted, and more efforts should be done to improve the re-
tering (HBF) [79,86], HE and HBF (HEHBF) [79], and Local Gray- search on DHV image enhancement.
level Information Histogram (LGIH) [150], etc. Among above al-
gorithms, according to the frequency of use, AHE and CLAHE are 4. Literature survey of DHV recognition methods
the most commonly used two algorithms for increasing the con-
trast. Particularly, most researchers treated contrast enhancement 4.1. Taxonomy of DHV recognition methods
as the necessary preprocessing step in DHV recognition. However,
this viewpoint is not entirely correct. Li et al. [89], Wang et al. Numerous methods for DHV recognition have been proposed.
[90] and Wang et al. [2] studied the specific influence of differ- Therefore, it is necessary to identify a taxonomy that can facilitate

6
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

the categorization of these methods. Two typical taxonomies have

been respectively presented by Huang et al. [91] and Wang et al.
[2].
In Huang’s taxonomy [91], according to the macroscopic at-
tributes of the venous network, the methods are roughly taxon-
omized into two classes: shape-based and texture-based. Shape
based methods extract vein shape as the features, such as the
skeleton, positions and angles of short straight vectors, end points
and crossing points, dominant points, vein minutiae and knuckle
shapes. Texture-based methods focus on texture variations within
the vein image, which include sparse texture descriptors such as
SIFT and its variations, and dense texture descriptors such as Lo-
cal Binary Patterns (LBP) and its variations. In Huang’s taxonomy
[91], based on how the feature of the vein pattern is computed, the
methods are categorized into global and local ones as well. Global
methods extract holistic features from the whole image. Typical
global methods include subspace learning-based methods, corre-
lation analysis-based methods, and Line segment Hausdorff Dis-
tance (LHD) based methods, etc. Local methods encode the de-
tailed characteristics within a specific region (a small part of the
whole area), including both the textural and geometrical descrip-
tions. More specifically, there are 4 types of methods in Huang’s Fig. 15. The relationship between our taxonomy and Huang’s and Wang’s tax-
taxonomy, i.e., global-shape, global-texture, local-shape, and local- onomies.
texture.
In Wang’s taxonomy [2], the representative DHV recognition
methods were taxonomized into four groups: global topological-
based methods, global quantification analysis methods, local based methods, (2) texture-based methods, (3) subspace learning-
geometric-based methods, and local invariant-feature-based meth- based methods, (4) direction coding-based methods, and (5) deep
ods. Global topological-based methods extract spatial vascular learning-based methods. Structure-based methods usually first per-
topological structure features for recognition such as line-like and form an image segmentation algorithm or line detection algorithm,
curve-like features. In addition, the geometrical parameters of ves- and then extract structure features for recognition such as lines,
sel such as radius, angle, length and minutiae coordinates are also skeletons, points, minutiae and graph. Thus, we further divide
treated as extra structure information. Global quantification anal- structure-based methods into three subtypes, i.e., line/skeleton-
ysis methods use matrix analysis tools for recognition, such as based methods, points/minutiae-based methods, and graph-based
subspace learning-based methods. Local geometric-based methods methods. Texture-based methods usually exploited sparse descrip-
first segment the vein network into a binarized one to get a high- tor, dense descriptor, or other texture representations, such as Ga-
quality vascular network distribution, then extract some geometric bor feature and wavelet feature, for DHV recognition. Thus, we fur-
features, such as positions and angles of short straight-line vectors, ther divide texture-based methods into three subtypes, i.e., Gabor
vein knuckle shapes, endpoints and crossing points, and dominant and wavelet-based methods, dense texture descriptor-based meth-
points. Local invariant-feature-based methods mainly refers to SIFT ods, and sparse texture descriptor-based methods. Dense descrip-
and its variations. tors, constructing local features pixel by pixel over the input im-
In our opinion, both Huang’s taxonomy and Wang’s taxon- age, have been successfully and widely used for biometrics. The
omy have some shortcomings. In Huang’s taxonomy, the term of most popular dense descriptors are LBP and Histogram of Oriented
“shape” is easily confused with “hand shape”. In addition, it is Gradients (HOG). It is well known that the most classical sparse
not very proper to treat subspace learning methods as the global descriptors are SIFT and its variations. Dense descriptors were also
texture-based methods, because subspace learning methods and called as local invariant-feature-based methods. The relationship
global texture-based methods are usually regarded as two differ- between our taxonomy and Huang’s and Wang’s taxonomies is
ent kinds of methods in image processing and pattern recognition. shown in Fig. 15. In our taxonomy, the type of structure-based
In Wang’s taxonomy, four groups of methods do not include dense methods is the same as the type of shape-based methods of
texture descriptors such as LBP and its variations, and those Gabor Huang’s taxonomy. In order to distinguish from hand shape-based
or wavelet-based texture methods have not been mentioned. For biometrics technology, we use the term of “structure” instead of
palmprint and vein recognition, direction coding-based methods “shape”. The methods based on line/skeleton and graph in our tax-
have been widely used, and have achieved excellent recognition onomy correspond to the global shape-based methods of Huang’s
performance. Generally, direction coding-based methods first ex- taxonomy. The methods based on points/minutiae in our taxonomy
tract the direction information of palmprint or vein, then, encode correspond to the local shape-based methods of Huang’s taxonomy,
the direction to spatial binary code, and match the binary codes by and correspond to the local geometric-based methods of Wang’s
hamming distance. In recent years, deep learning technology has taxonomy. Subspace learning methods in our taxonomy correspond
brought a breakthrough in recognition performance for many bio- to the global texture-based methods of Huang’s taxonomy, and are
metrics systems [160,161], and has been used for DHV recognition. the same as the global quantification analysis methods of Wang’s
However, both Huang’s taxonomy and Wang’s taxonomy do not taxonomy. In our taxonomy, the type of texture-based methods is
include direction coding-based methods and deep learning-based the same as the type of local texture-based methods of Huang’s
methods. taxonomy. Local invariant-feature-based methods in Wang’s taxon-
Based on Huang’s and Wang’s taxonomies and an analysis of omy correspond to sparse descriptors-based method of our tax-
the newest developments of DHV recognition, we propose a new onomy. Compared with Huang’s and Wang’s taxonomies, our tax-
taxonomy in this paper. In this new taxonomy, we divide exist- onomy is more comprehensive because it covers all kinds of DHV
ing DHV recognition methods into five types, i.e., (1) structure- recognition methods.

7
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Fig. 16. The flowchart of a general vein line/skeleton-based method

It should be noted that there is no unique taxonomy for these ate the skeleton after smoothing and thinning [17,28,34]. In the
algorithms. It is worth further investigating the design of a better third step, besides binary vein line pattern and skeleton, vein line
taxonomy in the future. can also be represented as other feature forms such as geomet-
ric moments [51,93] and multiresolution representations of fea-
4.2. Performance measures ture points [9], etc. In the last step, according to different feature
representations, different matching/classification algorithms were
It is well known that there are two recognition modes for a bio- exploited. For geometric moments-based feature representations,
metrics system, i.e., identification and verification. Generally, the some classical classifiers were exploited for classification such as
rank 1 identification rate is usually used as the performance mea- SVM [93], Artificial Neural Networks [51], and K-Nearest Neighbors
sure of identification. In this paper, for the purpose of typesetting [51], etc. For binary vein line pattern-based feature representation,
of some tables, we abbreviate “rank 1 identification rate” as “Rank the matching algorithms or classification algorithms include corre-
1”. For verification, Equal Error Rate (EER) is the most frequently lation matching [71,72], pixel-to-area matching [22], Hamming dis-
used measure to evaluate the performance of different verification tance [37], Sparse Representation based Classification (SRC) [30],
methods, which is the value when False Reject Rate (FRR) is equal and Collaborative Representation based Classification (CRC) [31],
to False Accept Rate (FAR). etc. For skeleton-based feature representation, the commonly used
matching algorithms include Hausdorff distance [34,109], correla-
tion matching [24], line segment Hausdorff distance [17], and Line
4.3. Structure-based methods for DHV recognition
Edge Map [34], etc. In order to improve recognition accuracy, the
strategy of image registration-based matching has been proposed.
In this subsection, we will review three subtypes of structure-
Gupta et al. [94] performed template-based matching after image
based methods one by one.
registration based on Fourier-Mellin transform. Cao et al. [28] used
The flowchart of a line/skeleton-based method is depicted in
iterative closest bifurcation points (ICBP) algorithm to register two
Fig. 16. In Table 4, we summarize typical line/skeleton-based meth-
vein patterns, then performed matching by the ratio of overlapping
ods and list their recognition accuracies. Generally, line/skeleton-
between a thinned image and a dilated image.
based methods have five main steps. In the first step, an image
Several researchers focused on extracting vein line patterns, but
enhancement algorithm is applied to input DHV ROI image to im-
did not perform matching experiments. Thus, we do not list these
prove the image quality. The main aim of the second step is to
methods in Table 4. Wang et al. [92] extracted the skeletons of
extract vein line or skeleton. In the third step, the vein line or
DHV from gray-scale FIR images by the watershed algorithm. Wang
skeleton feature representation is generated. In the last step, the
et al. [38] [82] proposed an adaptive thresholding method based
matching and classification is performed.
on local gradients for DHV segmentation, which can be treated
In the second step, two kinds of algorithms were often used
as a modified version of Sauvola and Niblack segmentation meth-
for vein line extraction. The first one is segmentation-based algo-
ods. Rossan et al. [84] discussed the segmentation results of sev-
rithm, the second one is edge/line detection-based algorithm. Be-
eral adaptive thresholding algorithms including mean-C, median-C,
cause vein line patterns are dark or bright lines, segmentation al-
and Wellner’s thresholding, and compared their recognition per-
gorithms can be used to separate the vein lines from background
formance. Kang et al. [95] proposed a skeleton extraction method,
to generate binary vein line pattern. Here, binary vein line pattern
which can directly extract topographic features of DHV without the
refers to line pattern without being thinned. A lot of fast and ef-
procedure of binarization processing. In [46], Kang et al. proposed
fective segmentation algorithms have been exploited including wa-
another method to segment the DHV line pattern. Recently, Fan
tershed transformation [9], local thresholding [17,33,34,71,72,93],
et al. [96] proposed a DHV segmentation method based on multi-
global thresholding [94], adaptive thresholding [30,31], or hybrid
ple features decomposition. Sun et al. [97] presented a binary filter
thresholding [51]. Some authors exploited edge/line detection algo-
for DHV line extraction, which has fast processing speed and can
rithms to extracted binary vein line pattern. Ferrer et al. [37] used
achieve similar results as the matched filter based on Gabor filter.
two 2D Gabor filter with zero average to filter DHV image to
The flowchart of a general minutiae/points-based method
obtain binary vein line pattern. Gupta et al. [94] and Cao et al.
is shown in Fig. 17. In Table 5, we summarize typical
[28] proposed to use multi-scale matched filtering to extract bi-
minutiae/points-based methods and list their recognition accu-
nary vein line pattern. Zhang et al. [22] extracted vein lines by
racies. Generally, minutiae/points-based methods have five main
modified finite radon transform (MFRAT), which is a well-known
steps. In the first three steps, skeleton of DHV should be extracted.
palmprint principal lines extraction method. Kauba et al. [24] ex-
In the fourth step, minutiae/points feature was extracted from the
tracted vein lines by the methods of maximum curvature and prin-
skeleton. In the last step, minutiae/points matching algorithm was
cipal curvature. For those skeleton based recognition methods, the
performed.
binary vein line pattern should be further processed to gener-

8
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 4
Summary of the subtype of structure-based methods: line/skeleton-based methods.

Segmentation or
line extraction Feature Matching or
References Database algorithm representation classification Performance Year
Rank 1 EER

Lin et al. NCU Modified version Multiresolution The mean absolute N/A 2.3% 2004
[9] of watershed representations of error (MAE)
transformation feature points of criterion
the vein patterns
(FPVPs)
Badawi TU Local thresholding Binary vein line The ratio of the N/A FAR=0.03%; 2006
et al. [33] pattern count of FRR=7.84%
overlapped white
pixels
Badawi TU Local thresholding Binary vein line Correlation in N/A 3% 2006
et al. [71] pattern frequency domain
Shahin TU Local thresholding Binary vein line Fast spatial N/A 0.25% 2007
et al. [72] pattern correlation
Wang et al. NTU (FIR) Local adaptive Extracting skeleton Line segment N/A 0% 2007
[17] thresholding after applying the Hausdorff
algorithm thinning algorithm distances (LHD)
NTU (NIR) N/A 0%
Wang et al. BIT Local thresholding Extracting skeleton Hausdorff distance 58% N/A 2008
[34] after applying the
thinning algorithm
Line Edge Map 66%
Ferrer et al. GPDS 2D Gabor Binary vein line Hamming distance N/A 1.42% 2009
[37] filter-based vein pattern
line extraction
∗
Li et al. 100 Local thresholding: Geometrical SVM classification 95.5% N/A 2010
[93] subjects; NiBlack algorithm moments extracted
500 images from vein skeleton
Li et al. NTU Nevatia-Babu edge Binary vein line Modified version 90% N/A 2012
[109] detection pattern of Hausdorff
distance
Gupta et al. IITK-PDV Multi-scale Binary vein line Image 98.71% 2.43% 2014
[94] matched filtering image registration-based
and global correlation
thresholding matching in spatial
domain
Cao et al. Multi-scale Extracting skeleton Image 93.8% N/A 2015
[28] MyPalmVein matched filtering after applying the registration-based
thinning algorithm matching
Castro- 38 Combining global Geometric Artificial Neural 86.36% N/A 2015
Ortega subjects; and local adaptive moments Networks
et al. [51] 912 images thresholding
K-Nearest 84.38%
Neighbors
Joardar JU-NIR-V1 Adaptive Binary vein line SRC 97.34% N/A 2015
et al. [30] thresholding pattern
Joardar JU-NIR-V2 Adaptive Binary vein line CRC 98.42% N/A 2016
et al. [31] thresholding pattern
Zhang HKPU ; Extracting lines by Binary vein line Pixel-to-area 97.27% N/A 2016
et al. [22] Multispec- modified finite pattern matching
tral radon transform
(MFRAT)
Kauba Extracting lines by Binary vein line Correlation N/A 0.124% 2018
et al. [24] PROTECTVein Maximum pattern matching
(IDS Curvature
Reflected
950)
Extracting lines by N/A 0.373%
Principal Curvature

Fig. 17. The flowchart of a general minutiae/points-based method.

9
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 5
Summary of the subtype of structure-based methods: minutiae/points-based methods.

Feature
References Database Skeleton extraction representation Matching Performance Year
Rank 1 EER

Ding et al. 48 Adaptive Ending points and Feature point 99.1% N/A 2005
[98] subjects; thresholding, and crossing points distance-based
240 images thinning matching
Wang et al. NTU Local thresholding, Minutiae Modified Hausdorff N/A 0% 2008
[18] skeletonization distance-based
and smoothing matching
Kumar IITD Mexican hat Minutiae Hierarchical N/A 1.77% 2009
et al. [10] operator matching
processing and
thinning
Soni et al. IITK Gaussian filtering, Minutiae Euclidean 99.26% N/A 2010
[108] and skeletonization distance-based
minutiae matching
Yuan et al. BJU Morphological Minutiae Two step 99.33% N/A 2010
[40] operations and minutiae-based
thinning matching
Hartung NTU (FIR) Multi-scale Location-based Mutual N/A 0.57% 2012
et al. [99] filtering, fast spectral minutiae information-based
marching (SML) minutiae matching
skeletonization
NTU (NIR) N/A 0.55%
Hartung NTU (FIR) Multi-scale Vein minutia MCC matching N/A 1.88% 2013
et al. [19] filtering, fast cylinder-codes
marching (V-MCC)
skeletonization
NTU (NIR) N/A 1.45%
Akram 120 Adaptive Ending and Mahalanobis 96.97% 3.03% 2014
et al. [110] subjects; thresholding, and bifurcation points distance-based
1200 thinning matching
images
Hu et al. 20 Niblack dynamic Reference point Relative distances 98% N/A 2014
[100] subjects; thresholding, and feature points and the angles
100 images smoothing and between the
thinning adjacent
connections
Lee et al. NDU A global/local Minutiae End-points-tree 99.64% 0.74% 2015
[101] thresholding, and (EP-tree) based
Stentiford thinning matching
algorithm
Chuang CNA Gaussian low-pass Minutiae Dynamic pattern 99.76% 0.63% 2017
et al. [102] filtering, global tree and Cosine
thresholding, and distance
thinning algorithm

Ding et al. [98] and Akram et al. [110] proposed points-based main steps. In the first three steps, skeleton of DHV should be
recognition methods, in which the ending points and crossing extracted. In the fourth step, graph representation was extracted
points were extracted for matching. Hu et al. [100]. used the in- from the skeleton. In the last step, graph-based matching algo-
tersection points and the endpoints selected as feature points, rithm was performed. Zhu et al. [103] proposed a method which
and then extracted the reference point from these feature points. treated the DHV skeleton as a graph. In this graph, the nodes
Some authors proposed to extract general minutiae feature for are the minutiae and the edges are the lines of skeleton. The
DHV recognition [10,18,40,101,108,102]. Hartung et al. [99] made graph edit distance was exploited for graph matching. Zhang et al.
a comprehensive analysis of DHV recognition based on spectral [104] called the graph presented in [103] as the basic graph model,
minutiae. Hartung et al. [19] also exploited a well-known finger- and they improved the basic graph model to the detailed graph
print recognition algorithm, minutia cylinder-codes (MCC), for DHV model by adding more vertices and embedding global texture fea-
recognition. tures. Later, Li et al. [105] built a generalized graph model named
In matching step, different matching strategies were exploited as Width Skeleton Model (WSM) for DHV recognition, which con-
such as feature point distance-based matching [98], modified siders both the topology structure of DHV and the width infor-
Hausdorff distance-based matching [18], Euclidean distance-based mation of vein line. Huang et al. [91] proposed to use Factorized
minutiae matching [108], mutual information-based matching [99], Graph Matching algorithm for DHV graph matching. Lajevardi et al.
MCC matching [19], mahalanobis distance-based matching [110], [106] proposed a graph-based method named as Biometric Graph
and cosine distance-based matching [102], etc. Particularly, Lee Matching (BGM). There are three main stages of the BGM algo-
et al. [101] exploited the end-points-tree (EP-tree) to accelerate the rithm including graph registration, error tolerant graph matching
minutiae matching procedure. Chuang et al. [102] used a dynamic and distance computation. Recently, Zhong et al. [14,107] further
pattern tree to speed up the minutiae matching. studied BGM-based DHV recognition under uncontrolled environ-
The flowchart of a Graph-based method is shown in Fig. 18. In ment. Liu et al. [162] studied the problem of partially occluded
Table 6, we summarize typical graph-based methods and list their DHV recognition using improved BGM.
recognition accuracies. Generally, Graph-based methods have five

10
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Fig. 18. The flowchart of a Graph-based method.

Table 6
Summary of the subtype of structure-based methods: graph-based methods.

Skeleton extraction Feature Matching and

References Database algorithm representation classification Performance Year
Rank 1 EER

Zhu et al. NCUT Local thresholding: Basic graph model Graph edit 88.33% N/A 2012
[103] NiBlack algorithm, distance
and thinning
Lajevardi NTU (FIR) Extracting skeleton Graph Biometric graph N/A 0% 2014
et al. [106] using maximum matching (BGM)
curvature
algorithm
NTU (NIR) N/A 0.68%
Li et al. NCUT Multi-scale Width Skeleton Line segment sets 99.31% N/A 2016
[105] hierarchical Model based matching
segmentation
based on Gaussian
matched filtering
Zhang NCUT Local thresholding: Detailed graph Graduated 99.22% N/A 2016
et al. [104] NiBlack algorithm, model Assignment based
thinning graph matching
Huang NCUT The Gaussian Graph Factorized Graph 92.75% N/A 2016
et al. [91] matched filter, and Matching (FGM)
global
thresholding, and
thinning
Zhong GPDS Curvature point Graph Biometric graph N/A 5.45% 2019
et al. algorithm-based matching (BGM)
[14,107] segmentation, and
thinning
XJTU N/A 5.34%
Liu et al. JLU Morphological Graph Biometric graph 98.09% 2.02% 2020
[162] operators matching (BGM)

4.4. Texture-based methods for DHV recognition and wavelet transform, in which LBP feature is extracted after
wavelet transformation. Wang et al. [39] proposed the method of
In this subsection, we will review three subtypes of texture- Partition Local Binary Pattern (PLBP) for DHV recognition, in which
based methods one by one. the LBP feature were extracted from the sub-images of DHV. Li
A lot of researchers focused on exploiting Gabor or wavelet fea- et al. [114] proposed an improved PLBP for DHV recognition by
tures for DHV recognition. In Table 7, we summarize typical Gabor combining structure information and multi-scale PLBP. Later, Wang
or wavelet texture-based methods and list their recognition accu- et al. [113] proposed a modified version of PLBP for DHV recogni-
racies. Wang et al. [34] used Gabor wavelet for feature extraction, tion by adding error correction coding (ECC) and feature weighting.
in which the multi-region mean Gabor magnitude was utilized as Wang et al. [2] developed a quality-specific DHV recognition sys-
the texture feature. Wang et al. [35] extracted texture feature of tem in which the DHV images were divided into two groups: high
DHV by wavelet decomposition. Kumar et al. [111] extracted tex- quality images and low quality images. For high quality image, the
ture features from FIR-based DHV image. They first partitioned the texture features were directly extracted by discriminative local bi-
ROI into non-overlapping blocks, and applied Gabor wavelets with nary pattern (DLBP), and for the low quality one, the feature ex-
four orientations to blocks. Then, the mean of the convolution re- traction was performed by contrast enhancement DLBP. Han et al.
sults on each block was used to construct final features. Raghaven- [123] proposed multi-scale Histogram of Oriented Gradient (MS-
dra et al. [112] extracted texture information using Gabor trans- HOG) descriptor, which has better recognition performance than
form, and then exploited the SRC for classification. original HOG descriptor for DHV recognition. Besides LBP, HOG,
As mentioned above, the most popular dense descriptors and their variations, a lot of researchers proposed other dense de-
are LBP and HOG. In Table 8, we summarize typical dense scriptors for DHV recognition. Trabelsi et al. [83] proposed Circular
descriptor-based methods and list their recognition accuracies. Difference and Statistical Directional Patterns (CDSDP), a rotation
Some authors used the basic LBP descriptor for DHV recognition invariant texture descriptor, to extract DHV features, which en-
[29,56,76,91,103]. In order to improve the recognition accuracy, a codes the local gradient direction information. Meng et al. [60] pro-
lot of variations of LBP and HOG have been proposed for DHV posed a descriptor named as the Histogram of the Local Gabor
recognition. Wang et al. [122] proposed a method combining LBP Phase XOR Pattern with a Second Identification (HLGPXP-SI) to

11
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 7
Summary of the subtype of texture-based methods: wavelet and Gabor texture-based methods.

Matching
Feature or classifi-
References Database extraction cation Performance Year
Rank 1 EER

Wang et al. BIT Multi- Euclidian 80% N/A 2008

[34] region distance,
mean 1NN
Gabor
magnitude
feature
Wang et al. TJU Multi- Euclidean 99.59% 1.96% 2008
[35] resolution distance,
wavelet KNN-SVM
analysis.
Kumar IITD Gabor Euclidean N/A FAR:0.1%; 2009
et al. [111] feature distance GAR:96%
TU Gabor SRC 88.72% N/A 2012
Raghavendra transform
et al. [112]

Table 8
Summary of the subtype of texture-based methods: dense descriptor-based methods.

Matching or
References Database Feature extraction classification Performance Year
Rank 1 EER

Zhao et al. [76] 15 Modified LBP Chi square distance 97.30% N/A 2008
subjects,
150 images
Wang et al. [39] NCUT Partition LBP Chi square distance 90.88% N/A 2010
(PLBP)
Wang et al. [122] NCUT Multi-scale LBP Euclidean distance 99.02% 2.067%, 2011
based on wavelet
Wang et al. [113] NCUT Coded and Chi square distance 98.83% N/A 2011
Weighted PLBP
Wei et al. [73] 100 Local XOR pattern Chi square distance N/A 2.17% 2011
subjects, based on Uniform
4000 Discrete Curvelet
images Transform (UDCT)
Zhu et al. [103] NCUT LBP Chi square distance 92.55% N/A 2012
Meng et al. [47] FU Local block pattern Consine distance N/A 0.88% 2013
Meng et al. [60] FU A Histogram of the Chi square distance N/A 0% 2013
Local Gabor Phase
XOR Pattern
(HLGPXP)
FU B N/A 3.83%
Meng et al. [115] FU Local Gabor Cosine distance N/A 0.53% 2014
ordinal measure
Wang et al. [29] NCUT LBP Euclidean distance 96.2% 0.034% 2015
Khan et al. [56] UM LBP Chi square distance 95% N/A 2015
Raghavendra et al. CUC LBP SRC N/A 7.719% 2015
[54]
Log Gabor Filters SRC N/A 0.767%
HOG SRC N/A 2.001%
Local Phase SRC N/A 3.364%
quantization (LPQ)
Local Binary SRC N/A 11.391%
Patterns Variance
(LBPV)
Kauba et al. [85] VeinPLUS LBP N/A N/A 29.7% 2015
Li et al. [114] NCUT Multi-scale PLBP Euclidean distance, 98.83% N/A 2015
(MPLBP) 1 NN
Trabelsi et al. [83] Bosphorus CDSDP ANN 99.8% 0.01% 2016
Han et al. [123] 40 Multi-scale Euclidean distance, 95.45% N/A 2016
subjects, Histogram of KNN
200 images Oriented Gradient
(MS-HOG)
Huang et al. [91] NCUT LBP Chi square distance 97.74% N/A 2016
Wang et al. [2] CUMT Discriminative LBP Chi square distance N/A 0.058% 2017
(DLBP)
Vairavel et al. NCUT WLD Chi square 98% N/A 2019
[116] distance, KNN
LBP 96% N/A
HOG 99% N/A

12
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 9
Summary of the subtype of texture-based methods: sparse descriptor-based methods.

Matching
Feature or classifi-
References Database extraction cation Performance Year
Rank 1 EER

Wang et al. NCUT SIFT SIFT-DM N/A 0.12% 2009

[36]
Wang et al. NCUT Multiple Euclidean 97.95% N/A 2012
[74] Keypoints distance
Sets + SIFT
Huang NCUT OGM+SIFT SIFT-based 99.02% N/A 2013
et al. [119] local
feature
matching
Huang NCUT Multi- SIFT-based 99.61% 0.80% 2015
et al. [3] source local
keypoint feature
+SIFT matching
Zhang NCUT Multi-task 99.66% N/A 2015
et al. [120] GDRKG + SIFT SRC
Huang NCUT CCKG+SIFT MtSRC 99.90% N/A 2016
et al. [121]
Poornima 10 SURF N/A 94% N/A 2016
et al. [117] subjects;
50 images

represent the DHV features. Meng et al. [47] presented a local DHV recognition, Zhang et al. [120] impoved SIFT matching frame-
block pattern for DHV recognition, in which the phase image was work from two aspects. In key point detection stage, the strat-
first extracted by a bank of Gabor filters and further encoded by egy of Gaussian Distribution based Random Key-point Genera-
the local XOR pattern (LXP). Meng et al. [115] also proposed a tion (GDRKG) was proposed, in which the Gaussian distribution-
descriptor named as local Gabor ordinal measure (OM) for DHV based model was used to generate random points. In matching
recognition. In this method, the DHV image is first filtered by Ga- stage, Multi-task Sparse Representation Classifier (MtSRC) was in-
bor wavelets, and then the magnitude and phase information ob- troduced to impove the classification performance. Later, Huang
tained by filtering is further encoded by ordinal filters. Wei et al. et al. [121] improved the work of [120] in key point detection
[73] applied Uniform Discrete Curvelet Transform (UDCT) on a DHV stage, and proposed Centroid-based Circular Key-point Grid (CCKG)
image, and then encoded the Curvelet coefficients phase variance for key point detection. Compared with CDRKG, experimental re-
by local XOR pattern for feature extraction. Kauba et al. [85] in- sults show that CCKG has better efficiency for key point detection,
vestigated the performance of several DHV recognition methods and can help improve the recognition accuracy.
including LBP, Maximum Curvature (MC), Adaptive Binarisation
(AB), Wide Line Detector (WLD), and Repeated Line Tracking (RLT). 4.5. Subspace learning-based methods
Raghavendra et al. [54] evaluated the performance of several tex-
ture descriptors including LBP, Histogram of Gradients (HOG), Lo- In Table 10, we summarize typical subspace learning meth-
cal Phase quantization, Steerable pyramids, Log Gabor Filters, and ods and list their recognition accuracies. Subspace learning-based
Local Binary Patterns Variance (LBPV). Vairavel et al. [116] investi- methods have different implementation strategies. The first one
gated the recognition performance of three classical dense descrip- is “image-as-vector”. In this strategy, algorithms require that the
tors for DHV recognition including LBP, HOG and Weber local de- 2-D image data must be reshaped into 1-D vector. Representa-
scriptor (WLD). tive “image-as-vector” algorithms include PCA, Linear Discriminant
Recently, sparse descriptors have been repeatedly studied Analysis (LDA), Independent Component Analysis (ICA), which have
for DHV recognition. In Table 9, we summarize typical sparse also been applied to DHV recognition [61,124,125,127]. The second
descriptor-based methods and list their recognition accuracies. one is the kernel method, which uses a linear classifier algorithm
Wang et al. [36] proposed to exploit the basic SIFT algorithm for to solve non-linear problems by mapping the original non-linear
DHV recognition. Poornima et al. [117] investigated DHV recogni- observations into a higher-dimensional space. Liu et al. [128] pre-
tion by using the Speeded Up Robust Features (SURF). Wang et al. sented a DHV recognition method based on Kernel PCA (KPCA) and
[74,82] proposed a DHV recognition method using SIFT based on Fisher Linear Discriminant (FLD). The third one is matrix and ten-
the strategy of multiple key-points sets, in which key-points de- sor embedding methods which extract a feature matrix using a
tected from all training images of each hand were merged. In this straightforward image projection technique. Hsu et al. [42] used a
way, the recognition accuracy can be effectively improved. In order modified version of 2-dimensional PCA, (2D)2 PCA, for DHV recog-
to find sufficient local features to improve the recognition perfor- nition. Furthermore, Hsu et al. [126] proposed the local feature-
mance of SIFT, Tang et al. [118] proposed a “multi-level key point based ensemble 2-directional 2-dimensional Linear Discriminant
detection” strategy, which used both Hessian-Laplace detector and Analysis (2D)2 LDA for DHV recognition. Zhu et al. [25] proposed
Harris-Laplace detector to detect key points. Huang et al. [119] pro- to use Two-Phase Test Samples Representation method (TPTSR) for
posed a method, which first extracts the DHV representation of DHV recognition. Zhang et al. [22] used (2D)2 PCA method for band
Oriented Gradient Maps (OGMs), then applies SIFT on OGMs for selection on the HKPU multispectral DHV database.
matching. Based on the works of [118] and [119], Huang et al.
[3] proposed a more complete SIFT matching framework, in which 4.6. Direction coding-based methods
the key points were detected by Harris–Laplace detector, Hessian–
Laplace detector and difference of Gaussian (DoG) on OGMs. For In Table 11, we summarize direction coding-based methods and
list their recognition accuracies. Neves et al. [26] used Orthogo-

13
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 10
Summary of subspace learning-based methods.

References Database Methodology Performance Year

Rank 1 EER

Khan et al. [124] 200 images PCA N/A 0% 2009

Liu et al. [61] 50 subjects ICA 99% N/A 2011
500 images
Hsu et al. [42] NDU Modified 99.53% N/A 2011
(2D)2 PCA
Hsu et al. [126] NDU 98.5543% N/A 2012
LFBE(2D)2 LDA
Liu et al. [127] 300 Wavelet 99.94% N/A 2012
subjects +GDICA2
1800
images
Liu et al. [128] 100 KPCA+FLD 97.63% N/A 2012
subjects
500 images
Khan et al. [125] UM LDA N/A 0% 2013
Zhu et al. [25] NUAA TPTSR 89.88% N/A 2015
HKPU 89.97% N/A
Zhang et al. [22] HKPU Mul- (2D)2 PCA 83.78% N/A 2016
tispectral

Table 11
Summary of typical direction coding methods.

References Database Methodology Performance Year

Rank 1 EER

Neves et al. ITS Orthogonal N/A 9% 2014

[26] Line
Ordinal
Features
(OLOF)
Lee et al. CNA VeinCode 99.27% 0.51% 2014
[49]
Hsu et al. NDU Gaussian 99.44% 0.95% 2015
[129] directional
binary
code
Lee et al. NDU Minimum 99.84% 0.54% 2016
[130] directional
code and
non-
orientation
code.
Zhang HKPU Mul- 96.45% N/A 2016
et al. [22] tispectral Competitive
Code

nal Line Ordinal Features (OLOF) for DHV recognition, which is 4.7. Deep leaning-based methods
an effective direction coding-based method proposed for palmprint
recognition. Lee et al. [49] proposed VeinCode for DHV recognition, In Table 12, we summarize deep learning-based methods and
in which adaptive Gabor filters were used to filter DHV images, list their recognition accuracies. Li et al. [12] investigated deep
and then the filter results were encoded to bit string. In match- learning-based methods on DHV recognition. Four popular CNN
ing stage, Hamming distance was used to calculate the similar- models were evaluated including Caffe Reference Net, Caffe Alex
ity between two VeinCodes. Hsu et al. [129] proposed a method Net, GoogLeNet, and VGG-16 Net. For DHV recognition, Wang et al.
called as Gaussian Directional Binary Code (GDBC). The mechanism [58] constructed a task-specific DNN model fine-tuned from the
of GDBC is similar to the method of VeinCode. The main differ- VGG-face for feature extraction and introduced large margin dis-
ence between them is that the binary code of GDBC is obtained by tribution machine (LDM) for classification. Based on the work of
Gaussian directional filtering. Lee et al. [130] proposed the method [58], in order to improve discriminative power of feature represen-
of Minimum Directional Filtering Response (MDFR) for DHV recog- tation, Wang et al. [131] further exploited selective convolutional
nition, in which a Gaussian directional filter bank was exploited features model (SCF) and introduced spatial pyramid pooling (SPP)
to filter the DHV image. For each pixel, the directional number of to obtain more robust feature representation. Zhong et al. [14] ex-
minimum response is selected as the direction feature, and is en- ploited Deep Hashing Network (DHN) for efficient DHV recog-
coded to bit string, which is denoted as Minimum Directional Code nition. Using DHN algorithm, DHV images were converted into
(MDC). Zhang et al. [22] applied Competitive Code to DHV recog- 128-bit binary codes. Then, the Hamming distances were used to
nition, which is a classical direction coding-method proposed for represent the similarity of two codes. In order to better evalu-
palmprint recognition. ate the proposed DHN-based algorithm, the authors in [14] com-
bined three DHV databases including NCUT, GPDS, and XJTU to

14
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 12
Summary of deep learning-based methods.

Deep feature extraction Matching algorithm

References Database modal or Classifier Performance Year
Rank 1 EER

Li et al. [12] NCUT VGG-16 KNN 99.31% N/A 2016

Wang et al. [58] CUMT A fine-tuned DNN Large margin 99.2% N/A 2018
model based on distribution
VGG-Net machine (LDM)
Wang et al. [131] CUMT Selective convolutional Linear SVM N/A 0.06% 2018
feature model (SCF)
and spatial pyramid
pooling (SPP)
Zhong et al. [14] NCUT+GPDS+XJTU Deep Hashing Network Hamming distance N/A 0.6% 2019
(DHN)
Wang et al. [15] CUMT A fine-tuned VGG with Softmax N/A 0.185% 2019
better augmented
datasets and model
compression algorithm
Wang et al. [158] CUMT Convolutional Large margin N/A 0.015% 2020
descriptors based on distribution
CNN, semantic feature machine (LDM)
selection and
high-level feature
encoding and
aggregation

form a larger database. Wang et al. [15] investigated two prob- complexity and need long matching time. Therefore, they are more
lems in deep learning-based method. First, for training data aug- suitable for verification, but not for identification.
mentation, they proposed a hierarchical generative adversarial net- Subspace learning-based methods have been widely studied,
work (HGAN) consisting of a CycleGAN and a constrained CNN. but they have an obvious shortcoming i.e., they are sensitive to
Second, in order to reduce the parameter redundancy in the over- illumination, translation and rotation variances even if these vari-
parameterized network, they proposed to use model compression ances are small. Thus, the recognition performance of these meth-
by way of filter pruning and low rank approximation. Wang et al. ods is not good in real applications. But, subspace learning-based
[158] designed a multi-weighted co-occurrence descriptor encod- methods can be used for dimensionality reduction of other meth-
ing (MWCDE) model for vein recognition. MWCDE generates con- ods, such as dense texture descriptor-based methods.
volutional descriptors based on CNN, then performs semantic fea- Direction coding-based methods are robust to illumination vari-
ture selection, and conducts high-level feature encoding and aggre- ance. However, their sensitivity to translation and rotation vari-
gation. Finally, LDM is used for classification. ances has not been well solved. Thus, for touch-based DHV recog-
nition, direction coding-based methods can achieve very good
4.8. Other methods recognition performance, because the translation and rotation vari-
ances in touch-based capturing mode is small. But, in touchless-
Tanaka et al. [149] and Gupta et al. [52] used the method of based capturing mode, their recognition performance will decline
Phase-Only-Correlation (POC) for DHV recognition. Poornima et al. due to large translation and rotation variances.
[117] proposed correlation filter-based methods for DHV recogni- Deep learning-based methods have achieved good recogni-
tion, in which the Minimum Average Correlation Energy (MACE) tion accuracy for face recognition, which is a challenging pat-
filters were exploited for recognition. tern recognition task. Thus, we think deep learning-based meth-
ods is also promising for DHV recognition. However, limited by the
4.9. Discussions on the recognition performance of different methods small amount of data in presented databases, the power of deep
learning-based methods has not been revealed for DHV recogni-
In the field of DHV recognition research, structure-based meth- tion.
ods were the earliest investigated methods because structure is the
basic feature of DHV. However, the recognition accuracy of them is 5. Information fusion schemes
not very satisfactory. The main reason is that the feature quality of
vein line or skeleton is easily influenced by image quality and il- It is well-known that using multi-modal biometrics based on
lumination. But as one of basic methods, structure-based methods information fusion is an effective strategy to improve the robust-
can be combined with other kinds of methods to form a multi- ness of biometric systems. For DHV recognition, multi-feature and
feature-based schemes. multimodal based recognitions are the two main forms for infor-
In texture-based methods, wavelet and Gabor-based methods mation fusion.
have limited discriminative power. Thus, the recognition accuracy
of them is not high. 5.1. Multi-feature-based DHV recognition schemes
Dense descriptors are very popular hand-crafted methods. They
can be easily designed, have achieved very high recognition accu- In Table 13, we summarize typical multi-feature-based DHV
racy, and have fast matching speed. Thus, these methods can be recognition schemes. It can be seen that the most of these fusion
applied to both application scenarios, verification and identifica- schemes are performed at score level. Yuksel et al. [45] proposed
tion. But the recognition accuracy of these methods may be easily to combine geometry and appearance features for DHV recogni-
influenced by image quality and other image variances. tion. In this method, the appearance features were extracted by ICA
In some literatures, sparse descriptors have achieved good and Non-negative matrix factorization (NMF) methods, and the ge-
recognition accuracy, but these methods have high computational ometry feature was represented by LEM. Afterwards, the decision-

15
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 13
Summary of typical multi-feature-based DHV recognition schemes.

References Database Fused features Fusion strategy Performance Year

Rank 1 EER

Yuksel et al. [45] Bosphorus (1) ICA, (2) NMF, (3) Decision-level fusion; 98.16% N/A 2010
LEM Score level fusion
Zhu et al. [103] NCUT (1) LBP, (2) Graph Score level fusion 96.67% N/A 2012
Huang et al. [91] NCUT (1) LBP, (2) Binary Score level fusion 99.21% N/A 2016
Coding (BC), (3) Graph
Zhang et al. [104] NCUT (1) Graph, (2) LBP, Score level fusion 99.22% N/A 2016
Gabor, and HOG
Zhang et al. [22] HKPU Mul- (1) Line feature Score level fusion 99.17% N/A 2016
tispectral extracted by MFRAT,
(2) Competitive Code

level fusion (Borda count technique, majority voting) and score- tion. In this system, the vein structure was extracted by a multi-
level fusion (product rule, sum rule and weighted sum rule) strate- scale matched filtering for DHV recognition, and finger length and
gies were used for information fusion. Zhu et al. [103] made use width were used for hand geometry recognition. The final fusion is
of both the texture and geometry features for DHV recognition, conducted at the score level. Chaudhary et al. [136] investigated a
in which texture feature was represented by LBP, and geometry multimodal scheme by the fusion of palmprint and DHV, in which
feature was represented by the graph constructed by the crossing gaussian membership function was used to extract features from
points and endpoints of the DHV pattern. Huang et al. [91] pre- both palmprint and DHV. Bharathi et al. [137] proposed a hand
sented a method via hierarchical combination of shape and texture vein-based multimodal biometric recognition system by fusing fin-
clues by score level fusion, which combined three kinds of features ger vein, palm vein and DHV, in which both the Shearlet transform
with different properties including LBP (local texture), Binary Cod- and SIFT were utilized to extract features. Lin et al. [55] presented
ing (local shape), and Graph (global shape). Zhang et al. [104] pro- a bimodal system by the fusion of palmprint and DHV patterns.
posed a holistic-based representation for DHV recognition, which Gopal et al. [138] presented a multi-modal person authentication
was denoted as the textured detailed graph model. In this method, system by fusing palmprint, DHV and inner knuckle-print. Gupta
global texture features represented as PCA coefficients of patches et al. [13] proposed a multimodal biometric system that fuses slap
centered as all the vertices were embedded to the graph, and lo- fingerprints, DHV and hand geometry for robust person authenti-
cal texture features, i.e., LBP, Gabor, and HOG, were also embedded cation. Zhong et al. [14] made an intensive study of the fusion of
to the graph. Zhang et al. [22] proposed a method by fusing line- DHV and palmprint. They proposed two fusion strategies. In the
based algorithm and direction coding-based algorithm, in which first strategy, the DHV feature extraction method was BGM, and
the line feature was extracted by MFRAT and direction coding was the palmprint feature extraction method was Deep Hashing Net-
represented by Competitive Code. work (DHN). The fusion was conducted at the score level. In the
second strategy, only DHN was used for feature extraction on both
DHV and palmprint modalities, in addition, the fusion was per-
5.2. Multi-modal based DHV recognition schemes formed on different levels including pixel level, feature level, score
level and decision level.
In Table 14, we summarize multi-modal-based DHV recogni- In Table 14, we also list the performance of DHV recognition
tion schemes. Ferrer et al. [132] developed a multimodal biomet- before multi-modal information fusion. It can be seen that multi-
rics system by fusing hand geometry, palmprint, finger texture and modal information fusion can obviously improve the recognition
DHV for person authentication. Kumar et al. [10] presented a mul- accuracy of the whole biometric system.
timodal person authentication system by combining the informa-
tion of DHV and knuckle shape. In DHV modality, minutiae tri-
angulation was extracted as the feature. In knuckle shape modal-
ity, geometrical features were extracted. After the matching scores 6. Commercial products
were generated in these two modalities, a weighted score level
combination was used to authenticate the individuals. Raghaven- Nowadays, in the market there are more and more biometrics
dra et al. [133] proposed a multi-modal system by the information commercial products based on hand traits and vein traits. For ex-
fusion of palmprint and DHV. Lu et al. [41] combined hand geome- ample, Fujitsu Ltd. provides a highly reliable biometric authenti-
try and DHV for person authentication based on the LDA algorithm. cation system based on palm vein pattern recognition technology,
Wu et al. [134] presented a multimodal hand vein recognition sys- which is named as PalmSecure. In the past decade, PalmSecure
tem, which combines DHV, palm vein and finger vein for person has been successfully used in a range of applications. In recent
authentication. Ramalho et al. [43] combined palm vein and DHV year, the users can find more finger vein-based biometrics com-
biometrics at the score level. Based on score level fusion, Bu et al. mercial products. Though DHV biometrics has been investigated
[44] developed a contactless multimodal system combining palm- for three decades, it has not been widely used. In the market, it
print, palm vein, DHV, finger vein and hand geometry. Using score seems that there is only one successful product of DHV recogni-
level fusion, Park et al. [48] established a multimodal hand bio- tion, which is the “VP-II X Hand Vascular Scanner” produced by
metric system based on hand geometry and DHV pattern. Trabelsi Techsphere Co. Ltd [156]. VP-II X Hand Vascular Scanner and its
et al. [135] developed a multimodal biometric system by fusing variant versions are shown in Fig. 19. It can be seen that VP-II X is
hand vein and finger vein. Gupta et al. [52] proposed to fuse DHV a touch-based biometrics product. The VP-II X Hand Vascular Scan-
and hand fist boundary for person authentication. In DHV pattern, ner has fast verification speed (0.4 sec/person) and excellent accu-
vein line and minutiae features were extracted and fused in the racy (FAR: 0.0 0 01%, FRR: 0.01%). The first version of VP-II product
feature level. Then, the DHV pattern and the hand boundary were has been released in 20 0 0, and its subsequent versions have been
further fused at the score level. Gupta et al. [57] proposed a sys- successfully deployed and proven in the market with several hun-
tem by combining DHV recognition and hand geometry recogni- dred customers around the world, over many years.

16
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

Table 14
Summary of typical multi-modal-based DHV recognition schemes.

DHV recognition
References DHV database Fused modalities Fusion strategy performance before fusion Performance after fusion Year
Rank 1 EER Rank 1 EER

Ferrer et al. GPDS (1) DHV, (2) Hand Score level fusion; N/A 1.79% N/A 0.01% 2009
[132] geometry, (3) Decision level
palmprint, (4) finger fusion
texture
Kumar IITD (1) DHV, (2) knuckle Score level fusion N/A 1.77% N/A 1.14% 2009
et al. [10] shape
TU (1) DHV, (2) palmprint Score level fusion N/A 12.5% N/A 2.2% 2010
Raghavendra
et al. [133]
Lu et al. HU (1) DHV, (2) hand Score level fusion N/A 1.18% N/A 0.48% 2010
[41] geometry
Wu et al. HIT (1) DHV, (2) palm Score level fusion N/A 0.8975% N/A 0.0223% 2010
[134] vein, (3) finger vein
Ramalho IUL (1) DHV, (2) palm vein, Score level fusion N/A 1.43% N/A 0% 2011
et al. [43]
Bu et al. HIT (1) DHV, (2) palmprint, Score level fusion N/A 0.4815% N/A 0.0138% 2011
[44] (3) palm vein, (4)
finger vein, (5) hand
geometry
Park et al. KU (1) DHV, (2) hand Score level fusion N/A 1.19% N/A 0.06% 2013
[48] geometry
Trabelsi Bosphorus (1) DHV, (2) hand vein Score level fusion N/A 1.1% N/A 0.004% 2013
et al. [135]
Gupta et al. IITK-1 (1) DHV, (2) hand Score level fusion 93.34% 8.45% 100% 2015
[52] boundary 0.0001418%
Gupta et al. IITK-PDV (1) DHV, (2) hand Score level fusion 99.33% 3.12% 99.34% 1.87% 2016
[57] geometry
Chaudhary Bosphorus (1) DHV, (2) palmprint Score level fusion 89% N/A 99.83% N/A 2016
et al. [136]
Bharathi TU (1) DHV, (2) finger Score level fusion N/A 0.1% N/A 0.04% 2015
et al. [137] vein, (3) palm vein
Lin et al. HHUT (1) DHV, (2) palmprint Image level fusion N/A FRR:8.53% 98.8% FRR:1.2% 2015
[55] FAR:3.30% FAR:1.56%
Gopal et al. Bosphorus (1) DHV, (2) Palmprint, Score level fusion N/A GAR:80.5% N/A GAR:99.6% 2016
[138] (3) inner knuckle-print ; FAR:1% ; FAR:1%
Gupta et al. IITK-2 (1) DHV, (2) Score level fusion 97.14% 3.23% 100% 0.72% 2018
[13] fingerprint, (3) hand
geometry
Zhong NCUT+GPDS+XJTU (1) DHV (BGM based Score level fusion N/A FRR:3.27% N/A FRR:0 2019
et al. [14] method), (2) Palmprint FAR:4.95% FAR:0.74%
(DHN-based method)
(1) DHV (DHN based Pixel level fusion N/A N/A N/A 0.660%
method), (2) Palmprint
(DHN-based method)
Feature level N/A N/A N/A 0%
fusion
Score level fusion N/A N/A N/A 0.305%
Decision level N/A N/A N/A FRR:0 %;
fusion FAR:0.74%

ever, a number of challenges are not sufficiently considered, which

often appear in real world application scenarios. Meanwhile, there
also exist several directions in future research. We introduce them
in the subsequent paragraphs.

7.1. Challenges on special cases in real applications

In general, DHV systems have not been extensively used. Ex-

isting DHV biometric systems are mainly built in controlled con-
ditions without considering the complex variations in practice.
Fig. 19. VP-II X Hand Vascular Scanner and its variant versions In fact, some changes in temperature, stretch, and fatigue may
degrade recognition performance. Actually, variations in the real
world are more than these, for instance, there are some special
7. Challenges and future directions cases of hand including tattoo, elderly people’s hands, and the
hand with long hair, as shown in Fig. 20. Even though these chal-
The potential of DHV biometrics has been demonstrated by cur- lenges may be encountered in real applications, they have not been
rent studies in the literature with promising performance. How- studied yet.

17
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

majority of DHV biometric datasets is less than 300 and it is re-

ally small compared to that of ones of other popular biometrics.
Furthermore, although there are dozens of databases, the public
ones are quite limited, which impedes direct comparison of differ-
ent methods. Therefore, to improve the DHV methods, it is impor-
tant to collect large-scale databases (large numbers of both sub-
jects and images per subject) with a good diversity. The common
variations should be covered, involving the provisional ones in illu-
mination, viewpoint, temperature, hand status as well as the per-
manent ones with tattoos by modifying ink, dyes and pigments in
dermis or scars by injury or surgery. At the same time, images
Fig. 20. Some special cases of hand. (a) tattoo; (b) elderly person’s hand; (c) the
need to be shot at multiple sessions of long-time lapses, with a
hand with long hair.
balanced distribution in terms of gender, ethnicity, age, etc.

7.2. Deep learning-based methods 7.5. Gender recognition

In biometric systems, recognition engines play the most funda- As one of important soft biometrics, gender recognition has
mental role. The methods are required to possess a strong ability to been studied for a long time. There are many works to recog-
distinguish vein patterns while presenting high robustness to the nize gender from face, fingerprint, iris, hand, and body. In the field
variations. Instead of the handcrafted features, those delivered by of DHV biometrics, Wang et al. [58] proposed different strategies
deep learning models, e.g., CNNs, can be regarded as a possible al- for gender recognition including fine-tuned DNN models, a two-
ternative. Although deep learning has proved very effective in im- layered unsupervised sparse feature learning model [146], and hi-
proving performance of computer vision tasks, they are seldom ap- erarchical transfer learning model [147]. Using these methods, the
plied in DHV recognition. In [12], the authors point out that some gender recognition rates of three methods can reach 91.6%, 98.2%,
pre-trained CNN features achieve promising results and more im- and 91.6%, respectively. From the works of [58,146,147], it can be
portantly, sophisticated pipelines in DHV recognition can be greatly seen that gender recognition based on DHV has been preliminar-
simplified. Recently, some advanced technologies in deep learning ily studied. Thus, in the future, more effort should be invested in
were used for biometrics. For example, neural architecture search promoting the research of this problem.
(NAS) technology has been successfully exploited for palmprint
and palm vein recognition, and has achieved promising recogni- 7.6. DHV recognition across different devices or different spectrums
tion accuracy [163]. In addition, Transformer is the latest trendy in
deep learning, which is an architecture that uses Attention to sig- In the field of biometrics, one of the key problems for real
nificantly improve the performance of deep learning. From 2020, applications is biometric sensor interoperability. In real applica-
the technology of Transformer began to be used for some visual tions, the biometrics data of the same individual may be cap-
tasks such as detection, classification, and segmentation, etc [164]. tured from different sensors or devices. That is, recognition is not
However, up to now, there has been no report on the application of always homogeneous. It is thus problematic if recognition is ap-
NAS and Transformer technologies to the DHV biometrics. There- plied across different systems, evidenced by poor performance re-
fore, more effort is needed to promote deep learning based DHV ported in [139]. Thus, a robust biometrics system must have a good
recognition. ability to perform heterogeneous DHV recognition. Wang et al.
[139,140,159] studied the problem of heterogeneous DHV recog-
nition across different devices. In [139], four methods including
7.3. The problem of biometrics aging based on DHV
deep learning, LBP, PCA and SIFT were used to evaluate the per-
formance of this task. The experimental results showed that the
In biometrics, the problem of aging has received wide atten-
performance of heterogeneous DHV recognition was very poor. To
tion. This problem refers to the gradual decrease in a system per-
investigate the problem of heterogeneous DHV recognition, Wang
formance caused by the changes suffered by the users’ trait over
et al. [140] proposed an improved SIFT algorithm, in which a seg-
time. That is to say, the biometrics systems should be robust to
mentation method based on gradient difference was firstly used
aging, in particular over a long period of time. The aging prob-
to segment the texture of veins, and then SIFT was used to ex-
lem has been investigated in other biometrics techniques. Sawant
tract and match the features. In this way, the recognition accuracy
et al. [153] made a survey on facial aging problem. Galbally et al.
was shown to reach 88%. Recently, Wang et al. [159] proposed a
[154] studied the aging problem in fingerprint biometrics. Fenker
method based on bit plane and block mutual information. In this
et al. [155] made an analysis on the template aging problem in iris
method, the vein line pattern was converted to eight-bit planes
biometrics.
to overcome the interference of brightness inside the higher bit
From Table 2 and 3, it can be seen that only 12 DHV image
planes and the interference of noise inside the lower bit planes.
databases were captured in multiple sessions with short intervals,
Then, the texture of each bit plane of dorsal hand vein was de-
e.g., one week to one month. Thus, the aging problem in DHV bio-
scribed by a block method and the mutual information between
metrics has not been studied.
blocks was calculated as texture features by three kinds of modes
to solve the problem of rotation and size. The accuracy of this
7.4. The construction of large-scale and public DHV image databases method for cross-device recognition can reach 93.3%.
Kauba et al. [24] investigated the problem of cross-spectrum
In the research on DHV recognition, the big challenge is the and cross-illumination DHV recognition. To do this, they created
lack of large-scale and public image databases. As we know, in the PROTECTVein database, which is a cross-spectrum and cross-
biometrics, data is crucial to evaluation of methods, and an ideal illumination DHV image databases. In the PROTECTVein database,
database is basically expected to be of a large scale, which is re- the DHV images of one hand were captured in both transmission
lated to number of subjects and number of samples per subject. mode and reflection mode, and two light sources were used in-
However, as shown in Table 2 and 3, the number of subjects in the cluding 850 nm LEDs and 950 nm LEDs in reflection mode. The

18
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

experimental results clearly showed that the cross-spectrum and the same time, more efforts should be conducted to solve those
cross-illumination DHV recognition has a low recognition perfor- challenging problems to improve the robustness of DHV biometrics
mance. systems. Thus, more high-quality research works can be expected.
In the future, in order to better solve the problem of hetero- In digital society, there are more and more application require-
geneous DHV recognition, the standardization of the device, espe- ments for biometrics. As an effective biometrics technique, DHV
cially imaging hardware, is a necessity where high-quality homo- recognition has a certain application potential. More and better
geneous DHV images can be stably obtained. DHV biometrics products should be launched to meet various ap-
plication requirements. Moreover, the products are expected to be-
7.7. Presentation attack detection and liveness detection for DHV come smaller and smaller so that they can be easily integrated into
recognition various portable devices, e.g., smart phones. Besides, configuration
solutions are needed for image acquisition without users’ cooper-
For real applications, improving the presentation attack detec- ation and even at a distance, in which case, the range of DHV ap-
tion (PAD) ability of a biometrics system is becoming increasingly plications can be substantially extended.
important. Pitil and Kanhangad et al. [141] investigated the vulner-
ability of the DHV biometric systems to display based presenta- 8. Conclusions
tion attacks. They collected a database consisting of 624 real and
312 spoof images from 52 subjects. Experimental results showed This paper provided a comprehensive survey on DHV biomet-
that DHV biometric systems are also vulnerable to spoofing at- rics. Particularly, the developmental history of DHV biometrics has
tacks. Later, they proposed a HOG based presentation attack detec- been presented, which can help the readers to better understand
tion in a DHV biometric system, which achieved an average error the progress of this technology. Existing DHV databases have been
rate of 0.16% and 0.8% in the same-device and the cross-device ex- listed. The DHV image quality assessment algorithms, and im-
periments, respectively [142]. age enhancement algorithms have been summarized, which are
Liveness detection is also a challenging task in biometrics. To very important issues in real applications. A new taxonomy for
detect the distinction between live and fake DHV, Wang et al. DHV recognition methods has been proposed, reflecting the lat-
[143–145] built a faked DHV database named as NCUT-LFDHV con- est developments in DHV biometrics. Furthermore, we systemat-
sisting of 100 printed dorsal hand vein images, 100 images of dor- ically reviewed existing recognition methods and information fu-
sal hand vein with a rubber thick glove and 100 images with a sion schemes, as well as discussing other important challenges
rubber thin glove. Using the NCUT-LFDHV database, Wang et al. and future directions in DHV biometrics research. In summary,
proposed to use Fourier spectral [143], AutoRegressive model [144], DHV recognition has become a mature biometric which is able to
and a combination of PCA and power spectrum estimation of the achieve a good recognition accuracy but there are still many im-
AR model together [145] for DHV liveness detection. portant challenges that have to be overcome.
In the future, more research should be conducted to improve
the performance of DHV presentation attack detection and liveness Declaration of Competing Interest
detection.
The authors declare that they have no known competing finan-
7.8. Biometric template protection for DHV images cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
In real applications, the server stores the biometric informa-
tion of users called templates. With the wide usage of biomet- Acknowledgments
rics technologies, the security of biometrics templates is becoming
more and more important. Sandhya et al. [157] made a systematic This work is partly supported by the grants of the National
literature review on biometric template protection. They divided Natural Science Foundation of China, Nos. 62076086, 61673157,
the schemes of biometric template protection into four types: 61972129, 61602540 and partly supported by the grants of
cancelable biometrics, biometric cryptosystems, hybrid methods, the Key Research and Development Program in Anhui Province
and Homomorphic Encryption based methods. For DHV biomet- 2020 04d07020 0 08, partly supported by the Open Research Fund of
rics template protection, few works have been done. Yildiz et al. the Beijing Key Laboratory of Big Data Technology for Food Safety
[148] discussed the DHV biometrics encryption issue, and pre- (Project No. BTBD-2021KF05), and partly supported by the Univer-
sented a Lorenz-like chaotic system to encrypt the DHV images. sity of Macau (File no. MYRG2018-0 0 053-FST).
Recently, Kauba et al. [165] found that it is possible to reconstruct
the gray-scale vascular images from their binary templates using References
a suitable CNN architecture, and the reconstructed gray-scale vas-
[1] G. Jaswal, A. Kaul, R. Nath, Knuckle print biometrics and fusion
cular images can be used for biometric recognition purposes. The schemes–overview, challenges, and solutions, ACM Comput. Surv. 49 (2)
work of [165] revealed that the security and privacy of even bi- (2016) 34.
nary templates can be compromised and that improved security [2] J. Wang, G. Wang, Quality-specific hand vein recognition system, IEEE Trans.
Inf. Forensics Secur. 12 (11) (2017) 2599–2610.
constructs are of vital importance in practical applications of DHV [3] D. Huang, Y. Tang, Y. Wang, L. Chen, Y. Wang, Hand-dorsa vein recognition by
biometrics. matching local features of multisource keypoints, IEEE Trans. Cybern. 45 (9)
(2015) 1823–1837.
[4] J. Kosmala, K. Saeed, Human identification by vascular patterns, Biometr. Kan-
7.9. Future directions
sei Eng. (2012) 67–87.
[5] J. Rice, Apparatus for the Identification of Individuals, Google Patents, 1987.
As we have mentioned above, the research on DHV biomet- [6] P. MacGregor, R. Welford, Veincheck: imaging for security and personnel
identification, Adv. Imaging. 6 (7) (1991) 52–56.
rics has achieved great progress, but there are many challenging
[7] J.M. Cross, C.L. Smith, Thermographic imaging of the subcutaneous vascular
problems which have not been well solved. In particular, the con- network of the back of the hand for biometric identification, in: Proc. IEEE
struction of large-scale databases with good diversity is very nec- 29th Annu. Int. Carnahan Conf. Security Technology, 1995, pp. 20–35.
essary and urgent. For DHV recognition, deep learning-based meth- [8] A. H. Choi, and C. N. Tran, Hand Vascular Pattern Technology, Handbook of
Biometrics, Springer, 2008, pp. 253-270.
ods should receive more attention. New theories and techniques [9] C.L. Lin, K.C. Fan, Biometric verification using thermal images of palm-dorsa
of deep learning should be timely used for DHV recognition. At vein patterns, IEEE Trans. Circuits Syst. Video Technol. 14 (2) (2004) 199–213.

19
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

[10] A. Kumar, K.V. Prathyusha, Personal authentication using hand vein triangula- national Conference on Frontier of Computer Science and Technology, 2010,
tion and knuckle shape, IEEE Trans. Image Process. 18 (9) (2009) 2127–2136. pp. 532–536.
[11] D. Huang, Y. Tang, Y. Wang, L. Chen, Y. Wang, Hand-dorsa vein recognition by [42] C.B. Hsu, S.S. Hao, J.C. Lee, Personal authentication through dorsal hand vein
matching local features of multisource keypoints, IEEE Trans. Cybern. 45 (9) patterns, Opt. Eng. 50 (8) (2011) 087201-1-087201-10.
(2015) 1823–1837. [43] M. Ramalho, P.L. Correia, L.D. Soares, Biometric identification through palm
[12] X. Li, D. Huang, and Y. Wang, Comparative Study of Deep Learning Methods and dorsal hand vein patterns, in: 2011 IEEE EUROCON - International Con-
on Dorsal Hand Vein Recognition, in: CCBR, 2016, pp. 296-306. ference on Computer as a Tool, 2011, pp. 1–4.
[13] P. Gupta, P. Gupta, Multi-biometric authentication system using slap finger- [44] W. Bu, Q. Zhao, X. Wu, Y. Tang, K. Wang, A novel contactless multimodal bio-
prints, palm dorsal vein and hand geometry, IEEE Trans. Indust. Electron. 65 metric system based on multiple hand features, in: 2011 International Con-
(12) (2018) 9777–9784. ference on Hand-Based Biometrics, 2011, pp. 289–294.
[14] D. Zhong, H. Shao, X. Du, A hand-based multi-biometrics via deep hashing [45] A. Yuksel, L. Akarun, B. Sankur, Hand vein biometry based on geometry and
network and biometric graph matching, IEEE Trans. Inf. Forensics Secur. 14 appearance methods, IET Comput. Vision 5 (6) (2011) 398–406.
(12) (2019) 3140–3150. [46] W. Kang, Vein pattern extraction based on vectorgrams of maximal in-
[15] G. Wang, C. Sun, A. Sowmya, Learning a compact vein discrimination model tra-neighbor difference, Pattern Recognit. Lett. 33 (14) (2012) 1916–1923.
with GANerated samples, IEEE Trans. Inf. Forensics Secur. (2019), doi:10.1109/ [47] Z. Meng, X. Gu, Hand vein recognition using local block pattern, Electron. Lett.
TIFS.2019.2924553. 49 (25) (2013) 1614–1615.
[16] R. Gade, T.B. Moeslund, Thermal cameras and applications: a survey, Mach. [48] G. Park, S. Kim, Hand biometric recognition based on fused hand geometry
Vis. Appl. 25 (1) (2014) 245–262. and vascular patterns, Sensors 13 (3) (2013) 2895–2910.
[17] L. Wang, G. Leedham, S.Y. Cho, Infrared imaging of hand vein patterns for [49] J.C. Lee, C.H. Lee, C.B. Hsu, P.Y. Kuei, K.C. Chang, Dorsal hand vein recognition
biometric purposes, IET Comput. Vision 1 (3–4) (2007) 113–122. based on 2D Gabor filters, Imaging Sci. J. 62 (3) (2014) 127–138.
[18] L. Wang, G. Leedham, D. Siu-Yeung Cho, Minutiae feature analysis for infrared [50] M. Faundez-Zanuy, J. Mekyska, X. Font-Aragonès, A new hand image database
hand vein pattern biometrics, Pattern Recognit. 41 (3) (2008) 920–929. simultaneously acquired in visible, near-infrared and thermal spectrums,
[19] D. Hartung, M. Tistarelli, and C. Busch, Vein Minutia Cylinder-Codes (V-MCC), Cogn. Comput. 6 (2) (2014) 230–240.
in: ICB, 2013, pp. 1-7. [51] R. Castro-Ortega, C. Toxqui-Quitl, G. Cristóbal, J.V. Marcos, A. Padilla-Vivanco,
[20] A. Kumar, M. Hanmandlu, V. K. Madasu, and B. C. Lovell, Biometric authenti- R.H. Pérez, Analysis of the hand vein pattern for people recognition, in: Proc.
cation based on infrared thermal hand vein patterns, in: 2009 Digital Image of SPIE, 2015, p. 95992Q.
Computing: Techniques and Applications, 2009, pp. 331-338. [52] P. Gupta, P. Gupta, Multi-modal fusion of palm-dorsa vein pattern for accurate
[21] S. Joardar, A. Chatterjee, S. Bandyopadhyay, U. Maulik, Multi-size patch based personal authentication, Knowl.-Based Syst. 81 (2015) 117–130.
collaborative representation for Palm Dorsa Vein Pattern recognition by en- [53] A. Gruschina, “VeinPLUS: A transillumination and reflection-based hand vein
hanced ensemble learning with modified interactive artificial bee colony al- database, arXiv preprint arXiv:1505.06769, 2015.
gorithm, Eng. Appl. Artif. Intell. 60 (2017) 151–163. [54] R. Raghavendra, J. Surbiryala, C. Busch, Hand dorsal vein recognition: sensor,
[22] D. Zhang, Z. Guo, Y. Gong, Dorsal hand recognition, Multispectral Biometrics algorithms and evaluation, in: 2015 IEEE International Conference on Imaging
(2016) 165–186. Systems and Techniques, 2015, pp. 16–18.
[23] K. Chen, D. Zhang, Band selection for improvement of dorsal hand recog- [55] C.L. Lin, S.H. Wang, H.Y. Cheng, K.C. Fan, W.L. Hsu, C.R. Lai, Bimodal biomet-
nition, in: 2011 International Conference on Hand-Based Biometrics, 2011, ric verification using the fusion of palmprint and infrared palm-dorsum vein
pp. 264–267. images, Sensors 15 (12) (2015) 31339–31361.
[24] C. Kauba, A. Uhl, Shedding light on the veins - reflected light or transillumi- [56] M.H.M. Khan, Representation of dorsal hand vein pattern using local binary
nation in hand-vein recognition, in: ICB 2018, 2018, pp. 283–290. patterns (LBP), in: International Conference on Codes, Cryptology, and Infor-
[25] Q. Zhu, Z. Zhang, N. Liu, H. Sun, Near infrared hand vein image acquisition mation Security, 2015, pp. 331–341.
and ROI extraction algorithm, Optik 126 (24) (2015) 5682–5687. [57] P. Gupta, S. Srivastava, P. Gupta, An accurate infrared hand geometry and vein
[26] J.R.G. Neves, P.L. Correia, Hand veins recognition system, in: 2014 Inter- pattern based authentication system, Knowl.-Based Syst. 103 (2016) 143–155.
national Conference on Computer Vision Theory and Applications, 2014, [58] J. Wang, G. Wang, M. Zhou, Bimodal vein data mining via cross-selected–
pp. 122–129. domain knowledge transfer, IEEE Trans. Inf. Forensics Secur. 13 (3) (2018)
[27] J. Wang, G. Wang, M. Li, K. Wang, H. Tian, Hand vein recognition based on 733–744.
improved template matching, Int. J. Bioautom. 18 (4) (2014) 337–348. [59] W. Damak, R.B. Trabelsi, M.A. Damak, D. Sellami, Dynamic ROI extraction
[28] J. Cao, M. Xu, W. Shi, Z. Yu, A. Salim, P. Kilgore, MyPalmVein: A palm method for hand vein images, IET Comput. Vision 12 (5) (2018) 586–595.
vein-based low-cost mobile identification system for wide age range, in: 2015 [60] Z. Meng, X. Gu, Palm-dorsal vein recognition method based on histogram of
17th International Conference on E-health Networking, Application & Ser- local gabor phase xor pattern with second identification, J. Signal Process.
vices, 2015, pp. 292–297. Syst. 73 (1) (2013) 101–107.
[29] Y. Wang, W. Xie, X. Yu, L.K. Shark, An automatic physical access control sys- [61] J. Liu, J.J. Cui, D.Y. Xue, X. Jia, Palm-dorsa vein recognition based on inde-
tem based on hand vein biometric identification, IEEE Trans. Consum. Elec- pendent principle component analysis, in: 2011 International Conference on
tron. 61 (3) (2015) 320–327. Image Analysis and Signal Processing, 2011, pp. 660–664.
[30] S. Joardar, A. Chatterjee, A. Rakshit, A real-time palm dorsa subcutaneous vein [62] Y. Zhao, M.Y. Sheng, Application and analysis on quantitative evaluation of
pattern recognition system using collaborative representation-based classifi- hand vein image quality, in: 2011 International Conference on Multimedia
cation, IEEE Trans. Instrum. Meas. 64 (4) (2015) 959–966. Technology, 2011, pp. 5749–5751.
[31] S. Joardar, A. Chatterjee, A. Rakshit, Real-time NIR imaging of palm dorsa sub- [63] D. Hartung, S. Martin, C. Busch, Quality estimation for vascular pattern recog-
cutaneous vein pattern based biometrics: An SRC based approach, IEEE In- nition, in: 2011 International Conference on Hand-Based Biometrics, 2011,
strum. Meas. Mag. 19 (2) (2016) 13–19. pp. 1–6.
[32] M. Yakno, J. Mohamad-Saleh, B.A. Rosdi, New technique for larger ROI ex- [64] X. Jia, J.J. Cui, D.Y. Xue, Near infrared vein image acquisition system based on
traction of hand vein images, in: 2015 International Conference on BioSignal image quality assessment, in: 2011 International Conference on Electronics,
Analysis, Processing and Systems, 2015, pp. 82–87. Communications and Control, 2011, pp. 922–925.
[33] A.M. Badawi, Hand vein biometric verification prototype: A testing perfor- [65] J.J. Cui, Q. Li, X. Jia, An image quality assessment algorithm for palm-dorsa
mance and patterns similarity, in: Proc. Int. Conf. Image Process. Comput. Vis. vein based on multi-feature fusion, Adv. Mater. Res. 508 (2012) 96–99.
Pattern Recognit, 2006, pp. 3–9. [66] Y.D. Wang, C.Y. Yang, Q.Y. Duan, Multiple indexes combination weighting for
[34] Z. Wang, B. Zhang, W. Chen, Y. Gao, A performance evaluation of shape and dorsal hand vein image quality assessment, in: International Conference on
texture based methods for vein recognition, in: Proc. Congr. Image Signal Pro- Machine Learning & Cybernetics, 2015, pp. 239–245.
cess, 2008, pp. 659–661. [67] L. Wang, G. Leedham, A thermal hand vein pattern verification system, in:
[35] Y. Wang, T. Liu, J. Jiang, A multi-resolution wavelet algorithm for hand vein International Conference on Pattern Recognition and Image Analysis, 2005,
pattern recognition, Chin. Optics Lett. 6 (9) (2008) 657–660. pp. 58–65.
[36] Y. Wang, D. Wang, T. Liu, X. Lib, Local SIFT analysis for hand vein pattern ver- [68] C.L. Lin, An approach to improve the quality of infrared images of vein-pat-
ification, in: Proc.2009 International Conference on Optical Instruments and terns, Sensors 11 (12) (2011) 11447–11463.
Technology: Optoelectronic Information Security, 2009 751204-1-751204-8. [69] J. Uriarte-Antonio, D. Hartung, J.E.S. Pascual, R. Sanchez-Reillo, Vascular bio-
[37] M. Ferrer, A. Morales, L. Ortega, Infrared hand dorsum images for identifica- metrics based on a minutiae extraction approach, in: 2011 Carnahan Confer-
tion, Electron. Lett. 45 (6) (2009) 306–308. ence on Security Technology, 2011, pp. 1–7.
[38] Y. Wang, H. Wang, Gradient based image segmentation for vein pattern, [70] P. Ramsoful, M.H.-M. Khan, Feature extraction techniques for dorsal hand vein
in: International Conference on Convergence Information Technology, 2009, pattern, in: Third International Conference on Innovative Computing Technol-
pp. 1614–1618. ogy, 2013, pp. 49–53.
[39] Y. Wang, K. Li, J. Cui, Hand-dorsa vein recognition based on partition local [71] A. Badawi, M. Shahin, Frequency domain spectral hand vein patterns au-
binary pattern, in: IEEE 10th International conference on signal processing thentication, in: Proc. of 3rd IEEE Cairo Int. Biomedical Engineering Conf.
proceedings, 2010, pp. 1671–1674. (CIBEC20 06), 20 06, pp. 1–5.
[40] X. Yuan, Biometric verification using hand vein-patterns, in: 2010 IEEE Inter- [72] M. Shahin, A. Badawi, M. Kamel, Biometric authentication using fast corre-
national Conference on Wireless Communications, Networking and Informa- lation of near infrared hand vein patterns, Int. J. Biomed. Sci. 2 (3) (2007)
tion Security, 2010, pp. 677–681. 141–148.
[41] W. Lu, X. Qi, C. Liu, Fusion of palm dorsal vein and hand geometry for per- [73] S. Wei, X. Gu, A method for hand vein recognition based on Curvelet Trans-
sonal identification based on linear discriminant analysis, in: 2010 Fifth Inter- form phase feature, in: Proceedings 2011 International Conference on Trans-
portation, Mechanical, and Electrical Engineering, 2011, pp. 1693–1696.

20
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

[74] Y. Wang, Y. Fan, W. Liao, K. Li, L.-K. Shark, M.R. Varley, Hand vein recognition under uncontrolled environment based on biometric graph matching, IET Bio-
based on multiple keypoints sets, in: ICB, 2012, pp. 367–371. metrics 8 (2) (2018) 159–167.
[75] S. Zhao, Y. Wang, Y. Wang, Extracting hand vein patterns from low-quality [108] M. Soni, S. Gupta, M. Rao, P. Gupta, A new vein pattern-based verification
images: a new biometric technique using low-cost devices, in: Fourth Inter- system, Int. J. Comput. Sci. Inf. Secur. 8 (1) (2010) 58–63.
national Conference on Image and Graphics, 2007, pp. 667–671. [109] F. Li, L. Wang, Y. He, Vein identification using line features, in: 2012 Spring
[76] S. Zhao, Y.D. Wang, Y.H. Wang, Biometric identification based on low-qual- Congress on Engineering and Technology, 2012, pp. 1–4.
ity hand vein pattern images, in: 2008 International Conference on Machine [110] M.U. Akram, H.M. Awan, A.A. Khan, Dorsal hand veins based person identi-
Learning and Cybernetics, 2018, pp. 1172–1177. fication, in: 4th International Conference on Image Processing Theory, Tools
[77] B. Li, X. Yang, Z. Chen, Study of fusion iterative enhancement algorithm of and Applications, 2014, pp. 1–6.
hand vein image based on wavelet transfor, in: 2012 Fifth International Sym- [111] A. Kumar, M. Hanmandlu, H.M. Gupta, Online biometric authentication using
posium on Computational Intelligence and Design, 2012, pp. 54–56. hand vein patterns, in: Proc. IEEE Symp. Comput. Intell. Secur. Defence Appl.,
[78] M. Yakno, J.M. Saleh, B.A. Rosdi, Low contrast hand vein image enhancement, 2009, pp. 1–7.
in: ICSIPA 2011, 2011, pp. 390–392. [112] R. Raghavendra, Sparse representation for accurate person recognition using
[79] R.D. Prasanna, P. Neelamegam, S. Sriram, N. Raju, Enhancement of vein pat- hand vein biometrics, in: 2012 Eighth International Conference on Intelligent
terns in hand image for biometric and biomedical application using various Information Hiding and Multimedia Signal Processing, 2012, pp. 41–44.
image enhancement techniques, Procedia Eng. 38 (2012) 1174–1185. [113] Y. Wang, K. Li, L.K. Shark, M.R. Varley, Hand-dorsa vein recognition based
[80] A. Djerouni, H. Hamada, A. Loukil, N. Berrached, Dorsal hand vein image con- on coded and weighted partition local binary patterns, in: 2011 International
trast enhancement techniques, Int. J. Comput. Sci. 11 (1) (2014) 137–142. Conference on Hand-Based Biometrics, 2011, pp. 253–257.
[81] S.K. Jiang, L. Fu, K. Bing, Z.Y. You, Y.X. Zong, Dorsal hand vein enhancement [114] K. Li, G. Zhang, Y. Wang, P. Wang, C. Ni, Hand-dorsa vein recognition based on
and fake vein detection, in: AITA 2017, 2017, pp. 283–287. improved partition local binary patterns, in: CCBR 2015, 2015, pp. 312–320.
[82] Y. Wang, K. Zhang, L.K. Shark, Personal identification based on multiple key- [115] Z. Meng, X. Gu, Hand vein identification using local Gabor ordinal measure, J.
point sets of dorsal hand vein images, IET Biometrics 3 (4) (2014) 234–245. Electron. Imaging 23 (5) (2014) 05300.
[83] R.B. Trabelsi, A.D. Masmoudi, D.S. Masmoudi, Hand vein recognition system [116] K. Vairavel, N. Ikram, S. Mekala, Performance analysis on feature extraction
with circular difference and statistical directional patterns based on an artifi- using dorsal hand vein image, Soft Comput. (2019) 1–10.
cial neural network, Multimed. Tools Appl. 75 (2) (2016) 687–707. [117] S. Poornima, F. Nasreen, A.D.S. Prakash, A. Raghuraman, Versatile and eco-
[84] I. Rossan, M.H.-M. Khan, Impact of changing parameters when preprocessing nomical acquisition setup for dorsa palm vein authentication, Procedia Com-
dorsal hand vein pattern, Procedia Comput. Sci. 32 (2014) 513–520. put. Sci. 50 (2015) 323–328.
[85] C. Kauba, A. Uhl, Robustness evaluation of hand vein recognition systems, [118] Y. Tang, D. Huang, Y. Wang, Hand-dorsa vein recognition based on multi-level
in: 2015 International Conference of the Biometrics Special Interest Group keypoint detection and local feature matching, in: Proceedings of the 21st
(BIOSIG), 2015, pp. 1–8. International Conference on Pattern Recognition, 2012, pp. 2837–2840.
[86] P. Chithra, A. Kalaivani, A Comparison of the Vein Patterns in Hand Images [119] D. Huang, Y. Tang, Y. Wang, L. Chen, Y. Wang, Hand vein recognition based
with other image enhancement techniques, Int. J. Emerg. Trends Technol. on oriented gradient maps and local feature matching, in: ACCV 2012, 2012,
Comput. Sci. (2016) 132–136. pp. 430–444.
[87] H.B. Wang, L. Tao, X.Y. Hu, Novel algorithm for hand vein recognition based [120] R. Zhang, D. Huang, Y. Wang, Y. Wang, Improving feature based dor-
on retinex method and sift feature analysis, in: Electrical Power Systems and sal hand vein recognition through Random Keypoint Generation and fine–
Computers, Springer, 2011, pp. 559–566. grained matching, in: 2015 International Conference on Biometrics, 2015,
[88] H.B. Wang, L. Tao, Novel algorithm for enhancement of hand vein images pp. 326–333.
based on adaptive filtering and retinex method, in: 2012 IEEE International [121] D. Huang, R. Zhang, Y. Yin, Y. Wang, Y. Wang, Local feature approach to dorsal
Conference on Information Science and Technology, 2012, pp. 857–860. hand vein recognition by Centroid-based Circular Key-point Grid and fine–
[89] F. Li, T. Zhang, Y. Liu, G. WANG, Hand-dorsa vein recognition based on scale grained matching, Image Vision Comput. 58 (2017) 266–277.
and contrast invariant feature matching, IEICE Trans. Inf. Syst. 100 (12) (2017) [122] Y.D. Wang, Q.Y. Yan, K.F. Li, Hand vein recognition based on multi-scale LBP
3054–3058. and wavelet, in: 2011 International Conference on Wavelet Analysis and Pat-
[90] G. Wang, J. Wang, SIFT based vein recognition models: analysis and improve- tern Recognition, 2011, pp. 214–218.
ment, Comput. Math. Methods Med. 2017 (2017). [123] T. Han, Z. Wang, X. Yang, Dorsal hand vein recognition based on Gabor mul-
[91] D. Huang, X. Zhu, Y. Wang, D. Zhang, Dorsal hand vein recognition via hier- ti-orientation fusion and multi-scale HOG features, in: Proceedings Volume
archical combination of texture and shape clues, Neurocomputing 214 (2016) 10024, Optics in Health Care and Biomedical Optics VII, 2016.
815–828. [124] M.H.-M. Khan, R. Subramanian, N.M. Khan, Low dimensional representation
[92] L. Wang, G. Leedham, A watershed algorithmic approach for gray-scale skele- of dorsal hand vein features using principle component analysis (PCA), World
tonization in thermal vein pattern biometrics, in: International Conference on Acad. Sci. Eng. Technol. 49 (2009) 1001–1007.
Computational and Information Science, 2006, pp. 935–942. [125] M.H.-M. Khan, N.A.M. Khan, Investigating linear discriminant analysis (LDA)
[93] X. Li, X. Liu, Z. Liu, A dorsal hand vein pattern recognition algorithm, on dorsal hand vein images, in: Third International Conference on Innovative
in: 2010 3rd International Congress on Image and Signal Processing, 2010, Computing Technology, 2013, pp. 54–59.
pp. 1723–1726. [126] C.-B. Hsu, J.-C. Lee, P.-Y. Kuei, K.-C. Chan, Combining local and global features
[94] P. Gupta, P. Gupta, A vein biometric based authentication system, in: Interna- based on the eigenspace for vein recognition, in: 2012 International Sym-
tional Conference on Information Systems Security, 2014, pp. 425–436. posium on Intelligent Signal Processing and Communications Systems, 2012,
[95] W.X. Kang, H.S. Li, F.Q. Deng, Direct gray-scale extraction of topographic fea- pp. 401–405.
tures for vein recognition, Sci. China Ser. F 53 (10) (2010) 2062–2074. [127] J. Liu, D. Xue, J. Cui, Palm-dorsa vein recognition based on general discrimi-
[96] J. Fan, J. Yang, C. Wu, D. Ai, H. Song, A. Hao, Y. Wang, Multiple features de- nant independent component analysis, Int. J. Digit. Content Technol. Appl. 6
composition for subcutaneous vein extraction and measurement, IEEE Access (17) (2012) 466–474.
6 (2018) 11265–11277. [128] J. Liu, D. Xue, J. Cui, X. Jia, Palm-dorsa vein recognition based on kernel prin-
[97] S. Sun, S. Li, Z. Guo, Binary Filter for Fast Vessel Pattern Extraction, Neural cipal component analysis and fisher linear discriminant, J. Comput. Inf. Syst.
Process. Lett. 49 (3) (2019) 979–993. 8 (4) (2012) 1545–1552.
[98] Y. Ding, D. Zhuang, K. Wang, A study of hand vein recognition method, [129] C.B. Hsu, J.C. Lee, S.J. Chuang, P.Y. Kuei, Gaussian directional pattern for dorsal
in: IEEE International Conference Mechatronics and Automation, 2005, hand vein recognition, Imaging Sci. J. 63 (1) (2015) 54–62.
pp. 2106–2110. [130] J.C. Lee, T.M. Lo, C.P. Chang, Dorsal hand vein recognition based on directional
[99] D. Hartung, M. Aastrup Olsen, H. Xu, H. Thanh Nguyen, C. Busch, Comprehen- filter bank, Signal Image Video Process. 10 (1) (2016) 145–152.
sive analysis of spectral minutiae for vein pattern recognition, IET Biometrics [131] J. Wang, Z. Pan, G. Wang, M. Li, Y. Li, Spatial pyramid pooling of selective con-
1 (1) (2012) 25–36. volutional features for vein recognition, IEEE Access 6 (2018) 28563–28572.
[100] Y.P. Hu, Z.Y. Wang, X.P. Yang, Y.M. Xue, Hand vein recognition based on the [132] M.A. Ferrer, A. Morales, C.M. Travieso, J.B. Alonso, Combining hand biometric
connection lines of reference point and feature point, Infrared Phys. Technol. traits for personal identification, in: 43rd International Carnahan Conference
62 (2014) 110–114. on Security Technology, 2009, pp. 155–159.
[101] J.C. Lee, Dorsal hand vein recognition based on EP-tree, in: 14th IAPR Inter- [133] R. Raghavendra, M. Imran, A. Rao, G.Hemantha Kumar, Multimodal biomet-
national Conference on Machine Vision Applications, 2015, pp. 402–405. rics: Analysis of hand vein & palmprint combination used for person verifica-
[102] S.J. Chuang, Vein recognition based on minutiae features in the dorsal venous tion, in: 3rd International Conference on Emerging Trends in Engineering and
network of the hand, Signal Image Video Process. 12 (3) (2017) 573–581. Technology, 2010, pp. 526–530.
[103] X. Zhu, D. Huang, Hand dorsal vein recognition based on hierarchically struc- [134] X. Wu, E. Gao, Y. Tang, K. Wang, A novel biometric system based on hand
tured texture and geometry features, in: CCBR 2012, 2012, pp. 157–164. vein, in: 2010 Fifth International Conference on Frontier of Computer Science
[104] R. Zhang, D. Huang, Y. Wang, Textured detailed graph model for dorsal hand and Technology, 2010, pp. 522–526.
vein recognition: a holistic approach, in: ICB 2016, 2016, 2016, pp. 1–7. [135] R.B. Trabelsi, A.D. Masmoudi, D.S. Masmoudi, A new multimodal biometric
[105] X. Li, D. Huang, R. Zhang, Y. Wang, X. Xie, Hand dorsal vein recognition by system based on finger vein and hand vein recognition, Int. J. Eng. Technol. 4
matching Width Skeleton Models, in: 2016 IEEE International Conference on (2013) 3175.
Image Processing, 2016, pp. 3146–3150. [136] G. Chaudhary, S. Srivastava, S. Bhardwaj, Multi-level fusion of palmprint and
[106] S.M. Lajevardi, A. Arakala, S. Davis, K.J. Horadam, Hand vein authentication dorsal hand vein, Adv. Intell. Syst. Comput. (2016) 321–330.
using biometric graph matching, IET Biometrics 3 (4) (2014) 302–313. [137] S. Bharathi, R. Sudhakar, V.E. Balas, Hand vein-based multimodal biometric
[107] D. Zhong, H. Shao, S. Liu, Towards application of dorsal hand vein recognition recognition, Acta Polytech. Hungarica 12 (6) (2015) 213–229.

21
W. Jia, W. Xia, B. Zhang et al. Pattern Recognition 120 (2021) 108122

[138] S.Srivastava Gopal, S. Bhardwaj, S. Bhargava, Fusion of palm-phalanges print [165] C. Kauba, S. Kirchgasser, V. Mirjalili, A. Uhl, A. Ross, Inverse biometrics: gen-
with palmprint and dorsal hand vein, Appl. Soft Comput. J. 47 (2016) 12–20. erating vascular images from binary templates, IEEE Trans. Biometr. Behav.
[139] Y. Wang, L. Xu, Study of heterogeneous dorsal hand vein recognition based Identity Sci. (2021), doi:10.1109/TBIOM.2021.3073666.
on multi-device, in: CCBR 2015, 2015, pp. 287–296.
[140] Y. Wang, X. Zheng, Cross-device hand vein recognition based on improved Wei Jia received the B.Sc. degree in informatics from Central China Normal Uni-
SIFT, Int. J. Wavel. Multiresol. Inf. Process. 16 (2) (2018) 1840010-1-17. versity, Wuhan, China, in 1998, the M.Sc. degree in computer science from Hefei
[141] I. Patil, S. Bhilare, V. Kanhangad, Assessing vulnerability of dorsal hand-vein University of Technology, Hefei, China, in 2004, and the Ph.D. degree in pattern
verification system to spoofing attacks using smartphone camera, in: 2016 recognition and intelligence system from University of Science and Technology of
IEEE International Conference on Identity, Security and Behavior Analysis, China, Hefei, China, in 2008. He has been a research associate professor in Hefei
2016, pp. 1–6. Institutes of Physical Science, Chinese Academy of Science from 2008 to 2016. He is
[142] S. Bhilare, V. Kanhangad, N. Chaudhari, Histogram of oriented gradients based currently an associate professor in Key Laboratory of Knowledge Engineering with
presentation attack detection in dorsal hand-vein biometric system, in: 2017 Big Data, Ministry of Education, and in School of Computer Science and Information
Fifteenth IAPR International Conference on Machine Vision Applications, 2017, Engineering, Hefei University of Technology. His research interests include computer
pp. 39–42. vision, biometrics, pattern recognition, image processing and machine learning.
[143] Y. Wang, Z. Zhao, Liveness detection of dorsal hand vein based on the analysis
of Fourier spectral, in: Biometric Recognition, Springer, 2013, pp. 322–329.
[144] Y. Wang, Q. Qi, K. Li, Liveness detection of dorsal hand vein based on Au- Wei Xia received the B.Sc. degree in computer science from Anhui University of
toRegressive model, in: 2014 IEEE Computers, Communications and IT Appli- Science and Technology, Huainan, China, in 2018. He is master student in School of
cations Conference, 2014, pp. 206–210. Computer Science and Information Engineering, Hefei University of Technology. His
[145] Y. Wang, D. Zhang, Q. Qi, Liveness detection for dorsal hand vein recognition, research interests include biometrics, pattern recognition and image processing.
Person. Ubiquit. Comput. 20 (3) (2016) 447–455.
[146] J. Wang, G. Wang, Z. Pan, Gender attribute mining with hand-dorsa vein Bob Zhang is currently an Associate Professor in the Department of Computer and
image based on unsupervised sparse feature learning, IEICE Trans. Inf. Syst. Information Science at the University of Macau. He obtained his B.A. in Computer
E101D (1) (2018) 257–260. Science from York University in 2006, a M.A.Sc. in Information Systems Security
[147] G. Wang, J. Wang, Z. Pan, Bimodal vein recognition based on task-specific from Concordia University in 2007, and a Ph.D. in Electrical and Computer Engineer-
transfer learning, IEICE Trans. Inf. Syst. 100 (7) (2017) 1538–1541. ing from the University of Waterloo in 2011. After graduating from Waterloo he re-
[148] M.Z. Yildiz, O. Boyraz, E. Guleryuz, A. Akgul, I. Hussain, A novel encryp- mained with the Center for Pattern Recognition and Machine Intelligence, and later
tion method for dorsal hand vein images on a microcomputer, IEEE Access worked as a Post Doctoral Researcher in the Department of Electrical and Computer
7 (2019) 60850–60867. Engineering at Carnegie Mellon University. His research interests focus on medical
[149] T. Tanaka, N. Kubo, Biometric authentication by hand vein patterns, in: SICE biometrics, pattern recognition, and image processing.
2004 Annual Conference, 2004, pp. 249–253.
[150] J. Wang, G. Wang, M. Li, W. Du, W. Yu, Hand vein images enhancement based Yang Zhao received the B.Eng. and the Ph.D. degree from department of automa-
on local gray-level information histogram, Int. J. Bioautom. 19 (2) (2015). tion, University of Science and Technology of China, in 2008 and 2013. From 2013
[151] Z. Peng, Z. Wang, X. Yang, Hand vein image enhancement based on phase to 2015, he was a Postdoctoral researcher at the School of Electronic and Computer
congruency, in: Proc. SPIE 10462, 2017. Engineering, Peking University Shenzhen Graduate School, China. Currently, he is an
[152] A. Uhl, C. Busch, S. Marcel, R. Veldhuis, Handbook of Vascular Biometrics, associate professor at the School of Computer Science and Information Engineering,
Springer, 2020. Hefei University of Technology, China. His research interests include image process-
[153] M.M. Sawant, K.M. Bhurchandi, Age invariant face recognition: a survey on ing and computer vision.
facial aging databases, techniques and effect of aging, Artif. Intell. Rev. 52
(2019) 981–1008.
Lunke Fei received the Ph.D. degree in computer science and technology from the
[154] J. Galbally, R. Haraksim, L. Beslay, A study of age and ageing in fingerprint
Harbin Institute of Technology, China, in 2016. Since 2017, he is an associate pro-
biometrics, IEEE Trans. Inf. Forensics Secur. 14 (5) (2019) 1351–1365.
fessor at the School of Computer Science and Technology, Guangdong University of
[155] S. P. Fenker and K. W. Bowyer, Analysis of template aging in iris biometrics,
Technology, Guangzhou, China. His research interests include pattern recognition,
in IEEE Computer Society Biometrics Workshop, June 17, 2012, pp. 1-7.
biometrics, image processing, and machine learning.
[156] http://www.vascularscanner.com/
[157] M. Sandhya, M.V.N.K. Prasad, Biometric template protection: a systematic lit-
erature review of approaches and modalities, in: Biometric Security and Pri- Wenxiong Kang received the M.S. degree from Northwestern Polytechnical Univer-
vacy: Opportunities & Challenges in The Big Data Era, R. Jiang, S. Al-maadeed sity, Xi’an, China, in 2003, and the Ph.D. degree from the South China University of
A. Bouridane, D. Crookes, A. Beghdadi, Springer, 2016, pp. 323–370. Technology, Guangzhou, China, in 2009. He is currently a Professor with the School
[158] G.Q. Wang, C.M. Sun, A. Sowmya, Multi-weighted co-occurrence descriptor of Automation Science and Engineering, South China University of Technology. His
encoding for vein recognition, IEEE Trans. Inf. Forensics Secur. 15 (2020) research interests include biometrics identification, image processing, pattern recog-
375–390. nition, and computer vision.
[159] Y.D. Wang, H. Cao, X.C. Jiang, Y.Y. Tang, Recognition of dorsal hand vein based
bit planes and block mutual information, Sensors 19 (17) (2019) 3718. Di Huang received the B.S. and M.S. degrees in computer science from Beihang Uni-
[160] W. Jia, J. Gao, W. Xia, Y. Zhao, H. Min, J.T. Lu, A performance evaluation of versity, Beijing, China, and the Ph.D. degree in computer science from the Ecole
classic convolutional neural networks for 2D and 3D palmprint and palm vein centrale de Lyon, Lyon, F́rance, in 2005, 2008, and 2011, respectively. He joined the
recognition, Int. J. Autom. Comput. 18 (1) (2021) 18–44. Laboratory of Intelligent Recognition and Image Processing, School of Computer Sci-
[161] S. Zhao, B. Zhang, Joint constrained least-square regression with deep con- ence and Engineering, Beihang University, as a Faculty Member. His current research
volutional feature for palmprint recognition, IEEE Trans. Syst. Man Cybern. interests include biometrics, in particular, on 2D/3D face analysis, image/video pro-
(2021), doi:10.1109/TSMC.2020.3003021. cessing, and pattern recognition.
[162] F. Liu, S. Jiang, B. Kang, T. Hou, A recognition system for partially occluded
dorsal hand vein using improved biometric graph matching, IEEE Access 8
Guodong Guo received the B.E. degree in automation from Tsinghua University, Bei-
(2020) 74525–74534.
jing, China, the Ph.D. degree in pattern recognition and intelligent control from Chi-
[163] W. Jia, W. Xia, Y. Zhao, H. Min, Y.X. Chen, 2D and 3D palmprint and palm vein
nese Academy of Sciences, Beijing, China, and the Ph.D. degree in computer science
recognition based on neural architecture search, Int. J. Autom. Comput. 18 (3)
from University of Wisconsin Madison, Madison, WI, USA. He is an Associate Pro-
(2021) 377–409.
fessor with the Department of Computer Science and Electrical Engineering, West
[164] K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y.
Virginia University (WVU), Morgantown, WV, USA. His research interests include
Xu, Z. Yang, Y. Zhang, D. Tao, A survey on visual transformer, arXiv preprint
computer vision, biometrics, machine learning, and multimedia.
arXiv:2012.12556, 2021.

22