The Palm Vein Graph

Feature Extraction and Matching

Arathi Arakala, Hao Hao, Stephen Davis and K. J. Horadam

School of Mathematical and Geospatial Sciences, RMIT University, Melbourne, Australia

Keywords:

Biometric Graph, Graph Matching, Authentication, Palm Vein.

Abstract:

We present a graphical representation for palm vein patterns for use as biometric identiﬁers. The palm vein

image captured from an infra red camera is converted into a spatial graph. After image enhancement and

binarisation, the palm vein features are extracted from the skeleton using a novel two stage spur removal

technique. The location of the features and the connections between them are used to deﬁne a Palm Vein

Graph. Palm vein graphs are compared using the Biometric Graph Matching (BGM) Algorithm. We propose

a graph registration algorithm that improves over existing state of the art algorithms for graph registration.

We introduce 10 graph topology-based measures for comparing palm vein graphs. Experiments are conducted

on a public palm vein database. One of the introduced measures, an edge-based similarity, gave a deﬁnite

improvement in matching accuracies over other published results on the same database, especially for samples

with only a small common overlap area due to displacement. In addition, when the edge-based measure was

combined with one of three other topological features, we demonstrate a further improvement in matching

accuracy.

1 INTRODUCTION

Palm vein biometrics refers to the use of the unique

vascular pattern in the human palm for authentication.

Palm veins have the following advantages shared with

most vascular biometrics:

1. Internal location, which makes it hard to to spoof

or covertly acquire from an individual.

2. In-built liveness detection in the capture process.

Palm-vein images are captured using infra-red (IR)

cameras. The palm vein images are created by the

property that infra red light is absorbed by the de-

oxygenated haemoglobin creating dark regions cor-

responding to the veins. The rest of the human tis-

sues and skin reﬂect the IR creating bright regions in

the image. This property ensures that a live subject is

available for biometric capture.

3. The capture is harmless and often contactless.

4. The capture process is non-intrusivewith respect to

an individual’s personal space and is therefore more

socially acceptable.

5. A very low failure to enrol (FTE) rate.

Palm vein has been identiﬁed as the easi-

est (Watanabe et al., 2005) vascular biometric to work

with as the palm area is free of hair and skin color

variations thereby reducing capture noise. Also the

palm vein pattern is very detailed and intricate and

coversa larger area than other vascular biometrics like

the dorsal vein, ﬁnger vein and wrist vein.

In this paper, we investigate the palm vein patterns

captured in the PUT database (Kabacinski and Kowal-

ski, 2011). The palm vein patterns are represented

as biometric graphs. Biometric graphs have been

used to represent ﬁngerprints (Horadam et al., 2011),

retina patterns (Lajevardi et al., 2013) and dorsal hand

vein patterns (Lajevardi et al., 2014). The palm vein

graphs are compared using Biometric Graph Match-

ing (Lajevardi et al., 2013). The main contributions

of this work are:

1. A novel image feature extraction process that cre-

ates a palm vein graph representing the vascular struc-

ture.

2. An improved registration algorithm that enhances

the speed and registration accuracy compared to the

algorithm in (Lajevardi et al., 2014).

3. Graph-based distance measures that effectively

separate genuine and imposter comparisons of palm

vein patterns.

295

Arakala A., Hao H., Davis S. and Horadam K..

The Palm Vein Graph - Feature Extraction and Matching.

DOI: 10.5220/0005239102950303

In Proceedings of the 1st International Conference on Information Systems Security and Privacy (ICISSP-2015), pages 295-303

ISBN: 978-989-758-081-9

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

2 IMAGE PREPROCESSING AND

GRAPH EXTRACTION

2.1 The Database

This research uses palm vein images captured by

researchers at the Poznan University of Technol-

ogy (PUT) (Kabacinski and Kowalski, 2011). The

database has vein pattern images from the left and

right palms of 50 individuals. Each individual gave

12 samples of each palm, totalling 24 samples. Each

set of 12 samples was acquired over 3 sessions

with 4 images each, at least one week apart. The

database was chosen for two main reasons - a large

number of samples per individual palm (12 images)

and 3 separate sessions, one week apart. Several

other databases exist in the literature (Kumar and

Prathyusha, 2009),(Shahin et al., 2007),(Wang et al.,

2007), but they all take samples within one session.

The across session comparison is valuable to under-

stand the natural variation that will occur in a practical

scenario where individuals leave after enrolment and

return after a period to be veriﬁed without remember-

ing the exact location of their hand at enrolment.

2.2 Image Processing and Feature

Extraction

The infra red capture process often generates noisy

images with low contrast between the vein pattern

and background. The challenges of feature extraction

in the PUT database include noise of wrinkles and

other skin features on the hand, non-uniform back-

ground brightness and palm displacement within and

across the sessions. In order to extract features from

the palm vein image, speciﬁc image processing steps

are required to enhance the vein pattern. We used

Discrete Fourier Transform based method, gradient-

based method and thresholding method (Kabacinski

and Kowalski, 2010; Wenxiong and Qiuxia, 2014;

Gaikwad and Narote, 2013) to extract the vein pat-

tern and found that they failed to provide the contin-

uous rich structure of palm vein in the PUT database.

This paper has modiﬁed and combined several noise

reduction and vein enhancement methods for the bet-

ter performance of palm vein graph extraction.

The automatic processing of graph feature extrac-

tion from palm vein images can be summarized in

the following steps - Image pre-processing and vein

enhancement; image binarisation; and skeleton and

graph feature extraction. It is important to identify

the Region of Interest (ROI) : the spatial region where

the vein patterns are most visible and useful. A well

deﬁned ROI will enable a good registration between

compared palm vein graphs and will hence improve

overall processing time. However, the PUT database

has shown large displacements in captured images,

especially between the sessions. In fact, the region

in one image may only be partially present in another.

In addition, if the ROI is too small, it limits the num-

ber of features that can be extracted from the image.

This paper selects the entire captured image area as

the ROI. Then, 40 pixels inwards from the border was

masked to eliminate possible misidentiﬁcation of the

vein.

Image pre-processing and vein enhancement aim

to remove the artifacts and increase the contrast be-

tween vein and background. The palm vein image

was ﬁrst converted to greyscale. To avoid amplify-

ing the noise in the image, especially in homogeneous

areas, contrast-limited adaptive histogram equaliza-

tion (Zuiderveld, 1994) was applied on a small win-

dow (8x8 pixels) to make image contrast uniform at

different regions. The noise in the image was re-

duced using an anisotropic diffusion ﬁlter (Perona and

Malik, 1990). Linear anisotropic diffusion performs

image smoothing and restoration that minimizes the

noise without removing signiﬁcant components of the

image. An example of the original image and vein-

enhanced image is shown in Figures 1(a) and 1(b).

Figure 1: This ﬁgure shows the stages that a palm vein im-

age goes through before feature extraction. (a) (top-left)

Original image; (b) (top-right) Palm vein enhanced image;

Skeleton extracted from binarised image.

The next step is image binarisation. Skeleton

extraction requires image binarization to distinguish

vessel and background. The process of binarization

was described previously in (Lajevardi et al., 2013).

Brieﬂy, a family of 2D matched ﬁlters was used to

search for vessel segments along all possible direc-

tions in frequency domain. The maximum responses

over all possible orientations were calculated for each

ICISSP2015-1stInternationalConferenceonInformationSystemsSecurityandPrivacy

296

pixel position. A threshold was chosen based on em-

pirical observation that the veins occupy around 15%

of the pixels in the image. The image was binarised

using this optimum threshold. Then the binarised im-

age was morphologically dilated to connect broken

pixels. Isolated objects, in which the connected com-

ponents were less than 200 pixels in area, were re-

moved. A binarized image is shown in Figure 1(c).

The ﬁnal step is skeleton and graph feature ex-

traction. The skeleton was extracted from the bi-

nary image by morphologically removing pixels on

the boundaries and keeping pixels connected, shown

in Figure 1(d). Graph features of vertices and edges

were then detected from the skeleton.

2.2.1 The Biometric Graph

The biometric graph is a spatial graph with features

in a biometric represented as vertices and the con-

nections between those features represented as edges.

Often the spatial coordinates of the features become

the spatial attributes of the corresponding vertices.

In a vascular biometric, the vein bifurcations and

crossovers represent the vertices and their locations in

the image represent their attributes. A pair of vertices

are connected by an edge if their features were linked

by a vein in the original vein pattern. The biometric

graph representation for retina (Lajevardi et al., 2013)

and handvein (Lajevardi et al., 2014) have been

shown to be extremely effective when compared to

other existing templates. Mathematically, a biometric

graph is represented as g = (V, E, µ, ν), where V

represents a set of vertices, E represents the set of

edges between the vertices, µ is the vertex labeling

function and ν is the edge labeling function. For the

palm vein graphs in this research, µ associates each

vertex with its unique two dimensional spatial coor-

dinate while ν associates each edge with attributes of

edge length and slope. The main advantages of using

a graphical representation is that it is concise and

modular. There are no limits to the vertex, edge or

graph attributes that can be added to the graph. Thus

additional features deemed important for representing

the biometric in the future can easily be added on the

existing graph representation as new attributes.

To obtain the palm vein graph after skeletonisa-

tion, a 3 by 3 window was applied on each pixel of the

skeleton to test if it was an end or a junction (crossing

point or branching point). All junction and end points

were marked as vertices of the graph. An edge is the

link between two vertices. Edges were detected by

neighborhood tracking along the skeleton. All possi-

ble branches were tracked from the junction vertex.

Palm vein graph features represent the vascular

Figure 2: This ﬁgure shows the extracted graph and the cor-

responding original image. Left ﬁgures are original images

and right ﬁgures are extracted graphs with features of ver-

tices (ﬁlled dots) and edges (line segments connecting a pair

of vertices). (a), the top panel and (b) the bottom panel,

show two palm vein images and corresponding graphs cap-

tured at different sessions from the same person. A large

displacement is noticeable between the two images.

pattern with limited vertices and edges. The undesir-

able spurs caused by noise and tiny veins have an im-

pact on the registration and matching process. There

were two stages to remove undesirable spurs. Firstly,

morphological operations were applied to remove un-

desirable short spurs on the skeleton. However, re-

moving spurs also reduces the vertices on the ma-

jor skeleton that may alter the shape of the skeleton.

Therefore, the threshold for morphologicaloperations

involved in deﬁning short spurs was set to 10 pixels to

ensure correct representation of the vein shape. Sec-

ondly, there were a number of short to medium spurs

in the skeleton due to palm principle lines and wrin-

kles. These lines and wrinkles vary in the captured

images at different sessions. Palm principal lines and

wrinkles can be considered as part of the biometric

and are difﬁcult to completely remove; our experi-

ments have shown that removing the short-medium

spurs improves the process of registration and match-

ing. The short-medium spurs were removed after the

detection of vertices and edges. We used the follow-

ing algorithm: 1) calculate the lengths of all branches.

2) identify the short-medium branches that are not

linked to the end vertices ( these branches should not

be removed) 3) the short-medium branches, that are

linked to end vertices and whose lengths are less than

40 pixels are removed. The connecting vertex was

kept to form the correct vein shape while the edge and

ThePalmVeinGraph-FeatureExtractionandMatching

297

Figure 3: Average Degree Distribution of Palm Vein

Graphs.

the vertex at the end of the edge was removed.

Examples of palm vein graphs from an individual

captured at two different sessions are shown in the

right panels of Figure 2(a) and Figure 2(b).

The Palm vein graphs from the PUT database have

an average of 101 vertices and 112 edges. The aver-

age edge to vertex ratio is 1.11 and the average size

of the largest component is 91. The average size of

the largest component is almost equal to the aver-

age size of the palm vein graph. This indicates that

the image processing can retrieve most of the features

and connecting veins in a given image. Smaller dis-

connected components exist in a palm vein graph be-

cause the quality of the image in certain parts is not

good enough to retrieve the connecting veins. Fig-

ure 3 shows the degree distribution of the Palm Vein

Graphs.

3 BIOMETRIC GRAPH

MATCHING

Biometric Graph Matching (BGM) is a two-stage al-

gorithm for the error-tolerant comparison of biomet-

ric graphs and is described in Lajevardi et al (Laje-

vardi et al., 2013). The ﬁrst stage is registration and

the second is error tolerant graph matching. Regis-

tration is the process of bringing a pair of compared

graphs into the same reference frame. The Iterative

Closest Point (ICP) algorithm (Chen et al., 2009) and

the Modiﬁed Hausdorff Distance (MHD) (Dubuisson

and Jain, 1994) are algorithms commonly used to reg-

ister point clouds. It has been demonstrated by Laje-

vardi et al (Lajevardi et al., 2014) that the registration

process in BGM is either as good as or does better

than these standard algorithms, especially when the

graphs are small. This paper uses a registration al-

gorithm that is a slight modiﬁcation of Algorithm1 in

Lajevardi et al (Lajevardi et al., 2013). The modiﬁed

algorithm is denoted here as Algorithm 2 (see com-

parison results in Table 1). The registration process

in Algorithm 1 translates and rotates a pair of graphs

based on a corresponding pair of edges that deﬁne the

coordinate system. Every pair of edges with one edge

from each compared graph, is given a score based on

a function of their edge attributes of length and slope.

The edge pairs are ranked on this score. 2% of the

top ranked edges are then shortlisted for the registra-

tion process. For each of the shortlisted edge pairs,

the two graphs are translated and rotated to make the

start vertex of each edge the origin and the end ver-

tex indicate the positive x-direction. The vertex labels

of both the translated and rotated graphs are recalcu-

lated based on the new orientation. The Euclidean

distance between the vertex labels, which in this case

are the cartesian coordinates of the vertex positions,

are compared to count the number of vertex pairs that

lie within a tolerance ε of each other. The edge pair

that gets the largest number of vertex pairs within ε is

chosen as the edge pair to register the graphs and the

corresponding transform (translation and rotation) is

taken as the correct transform to achieve the best reg-

istration.

On implementing Algorithm 1 for registering the

palm vein graphs, we found one drawback that caused

several genuine palm vein samples to be misaligned.

Most edge pairs in the palm vein graph were short and

shorted pairs often scored a high rank compared to

longer pairs. This caused longer pairs that gave a bet-

ter registration to not appear on the top 2% shortlist.

To overcome this, we modiﬁed Algorithm1 to split the

set of edge pairs into long and short edge pairs. The

mean of the medians of the edge lengths in the two

graphs was selected as the threshold. If both edges

of an edge pair had length greater than this threshold,

the edge pair was categorised as long. All other edge

pairs were labeled as short. The new shortlist com-

prised the top 2% of long edge pairs and the top 2%

of short edge pairs. This modiﬁcation ensured that

long edge pairs that potentially gave better alignment

could be included in the shortlist to get a better regis-

tration of the graphs. This is called as Algorithm 2 in

this paper.

Once a pair of graphs is registered, the graph

matching process gives a numerical measure of the

distance between the pair of graphs by using the

topological properties of a graph edit path-based in-

duced subgraph. The error tolerant graph match-

ing is adapted from that proposed by Riesen and

Bunke (Riesen and Bunke, 2009). Their proposed al-

gorithm computes a graph edit distance and graph edit

path between a pair of compared graphs by treating

the two graphs as two vertex sets of a bipartite graph.

ICISSP2015-1stInternationalConferenceonInformationSystemsSecurityandPrivacy

298

Every primary vertex is mapped to one secondary ver-

tex using the Hungarian optimisation algorithm with

a cost function based on the vertex attributes. In this

paper, the cost function between a pair of vertices is

the Euclidean distance between the vertex labels and

is described in (Lajevardi et al., 2013).

The graph matching process is simpliﬁed to a set

of edit operations to convert the ﬁrst graph (source

graph) to the second graph (destination graph). The

edit operations allowed are insertion, deletion and

substitution of vertices. Each vertex in the source

graph has a ﬁxed cost associated with deleting it

(D(n) ×α) where D(n) is the degree of the vertex n

and α is called the insertion and deletion cost and is

a non zero integer. If n is a vertex in the destination

graph, the cost of inserting it is also D(n) ×α. In this

research we keep the insertion cost and deletion cost

the same (α), however, it is possible to have different

costs. Once the Hungarian algorithm computes the

cheapest edit path to convert the source graph to the

destination graph, this edit path is used to build the

maximum common subgraph (mcs). The vertices of

the destination path could be associated with two op-

erations from the edit path - substitutions of vertices

from the source graph and insertions. The mcs is a

vertex induced subgraph of the destination graph. In

particular, the mcs will comprise only those vertices

of the destination graph that were associated with sub-

stitutions in the cheapest edit path. An edge will exist

in the mcs if the edge existed in corresponding ver-

tex pairs of both the source and destination graphs.

Figure 4 shows the mcses from a genuine graph com-

parison and an imposter graph comparison. Note the

difference in topology of the two types of mcses. Mc-

ses from a genuine comparison are observed to have

more vertices and more edges with larger connected

components than the mcses from an imposter compar-

ison. It is this difference in topology that will be used

to deﬁne similarity scores to distinguish genuine and

imposter comparisons.

4 GRAPH FEATURES

The topological features of the mcs are used to

differentiate the comparisons of samples from the

same palm and samples from different palms. Let

= (V

, E

, µ, ν) and g

= (V

, E

, µ, ν) be the two

graphs compared using the BGM algorithm and g =

(V, E, µ, ν) be the mcs of the graphs. Let C1 and C2

be the vertex counts of the largest and second largest

components, c

and c

in g.

Similarity scores based on the graph features are

deﬁned as follows:

Figure 4: Examples of registrations of a pair of graphs taken

from the (a) (top - left) the same palm at different sessions

and (c) (bottom-left) from different palms. The correspond-

ing maximum common subgraphs are in (b) (top-right) and

(d) (bottom-right). In Figure (a) and (c) the edge pair on

which the graphs register is shown by the black circle.

1. Normalised vertex count, S

|V|

√

|×|V

2. Normalised edge count, S

|E|

√

|×|E

3. Edge to vertex ratio, S

|E|

|V|

4. Size of the largest component, S

√

|×|V

5. Size of the second largest component, S

√

|×|V

6. Ratio of the number of isolated vertices in the mcs

to the number of connected vertices, S

7. Total length of the edges in the largest component

in the mcs. Let L() be a function that takes a graph

as input and returns the sum of the lengths of the

edges in the graph. Then S

L(c

)

√

L(g

)×L(g

)

8. sum of the sizes of two largest components in the

mcs, S

c1c2

C1+C2

√

|×|V

9. Edge to vertex ratio of the largest component, S

10. Ratio of the number of vertices with degree

greater than 3 in the mcs, to the total number of

vertices in the mcs, S

We investigate if one or more of these features can

be combined to achievemaximum separation between

genuine and imposter scores of comparisons from the

palm vein database.

ThePalmVeinGraph-FeatureExtractionandMatching

299

Table 1: Comparison of Registration Algorithms Algo-

rithm1 (Lajevardi et al., 2013) and Algorithm 2, proposed

in this paper.

Algorithm 1 Algorithm 2

Best ε 6 6

Average time (sec) 15.99 25.94

EER (as %) 2 0

5 EXPERIMENT

The PUT database has for left and right hands re-

spectively, 600 palm vein images from 50 individuals,

giving 6600 genuine comparisons across all sessions,

1800 genuine comparisons within same sessions and

352800 imposter comparisons. To tune the param-

eters in the experiment, we chose a random sample

of 50 genuine comparisons and 50 imposter compar-

isons from all possible genuine and imposter compar-

isons. This set of 100 comparisons is called the train-

ing set and will be consistently used throughout all

the experiments described in this paper. The BGM

algorithm has two parameters to be tuned, the toler-

ance ε in the registration algorithm and the insertion

and deletion cost α in the graph matching algorithm.

The ﬁrst experiment was to compare the registration

algorithm by Lajevardi et al (Lajevardi et al., 2013)

with the improved algorithm, Algorithm 2, presented

in this paper. The graph pairs in the testing set were

aligned ﬁrst using Algorithm 1 and then Algorithm

2 and in each case, the number of vertex pairs that

lay within a tolerance ε of each other were counted.

A distance measure based on the Similarity score S

(Section 4, Score 1), given by d

min

= 1 −S

is com-

puted from the number of common vertex pairs and

the sizes of the two graphs compared. The d

min

val-

ues between genuine and imposter comparisons in the

training set are used to deﬁne score distributions to

compare the alignment performance of the two algo-

rithms. This experiment is run over range of ε values

to ﬁnd the ε value that gave the lowest EER for each

algorithm. The Equal Error Rates (EER) at the best ε

value for each algorithm and the average registration

times are presented in Table 1.

Setting ε = 6 and using Algorithm 2 for reg-

istration, the BGM algorithm was run on all the

comparisons from the palm vein database, excluding

the comparisons used in the training set. BGM was

run using a range of α values to determine the pa-

rameters that best separated the genuine and imposter

scores. It was found that α = 11 best compensated

for the variations within samples of the same hand

in determining the graph edit path when comparing

pairs of graphs.

The BGM algorithm was run on Left and Right hands

were tested separately. There were two distinct types

of experiments based on the type of genuine com-

parisons made. First, Across Session Genuines using

all the genuine comparisons including those across

sessions and the second Within Session Genuines,

with genuine comparisons taken only from within

the same session. In both types of experiments,

distances between a pair of graphs was measured

using each of the 10 topological measures listed in

Section 4. Table 2 shows the EER based on the the

10 topological features for across session and within

session experiments for left and right hands. The

Table 2 shows that S

does the best job of separating

the genuine comparisons from the imposters.

The next step was to determine if any combina-

tion of similarity scores could improve the perfor-

mance compared to using a single similarity measure.

To do this, ﬁrst the pairwise Spearman’s correlation

coefﬁcient between the 10 features was calculated.

We found that most of the topological features were

strongly correlated, with correlation coefﬁcients be-

tween 0.8 to 0.95. In fact only S

showed moder-

ate correlations ranging between 0.58 and 0.78 with

the other features. Nevertheless, in the absence of

perfect correlation between features, there is a poten-

tial that pairwise combination of S

with one of the

other features could produce better matching results.

To test this hypothesis, S

was combined with every

other feature to give 9 different pairings of topolog-

ical features. For every pair of features, the follow-

ing experiment was done. A Support Vector Machine

(SVM) was used to build a classiﬁer with a radial ba-

sis function (RBF) kernel that was tuned on the score

pairs from the training set to determine the best pa-

rameters for the RBF kernel. The remaining compar-

isons on the database were divided into 10 parts and a

10 fold test was conducted where the SVM classiﬁer

was trained on 9 parts of the data and tested on the

one other part. The false match rate (FMR), false non

match rate (FNMR) and total misclassiﬁcation error

(TE) were computed in every fold. The average over

10 folds was taken as the matching performance us-

ing the chosen pair of features. The results for the 9

pairings are shown in the Table 4.

6 RESULTS AND DISCUSSION

Table 1 shows that Algorithm 2 signiﬁcantly improves

the registration process over Algorithm1 by Lajevardi

et al (Lajevardi et al., 2013), evidenced by the 0%

ICISSP2015-1stInternationalConferenceonInformationSystemsSecurityandPrivacy

300

Table 2: Equal ErrorRate as % , when comparing genuine and imposter palm vein graphs using each of the 10 graph measures.

EER as %

Topological Across Session Within Session Across Session Within Session

Measure Genuines , Left Genuines, Left Genuines , Right Genuines, Right

4.18 2.02 4.17 2.13

2.71 1.11 2.55 1.23

5.32 2.9 5.27 3.42

6.06 3.91 6.63 4.39

7.33 5.49 7.76 5.58

8.43 5.66 8.25 5.93

5.64 3.5 6.59 4.19

c1c2

4.45 2.51 4.79 2.91

19.22 19.96 13.04 12.67

13.33 9.9 11.5 8.57

EER for Algorithm 2. This demonstrates that for palm

vein graphs, a better registration could be achieved by

considering longer edge pairs which represent longer

vein segments that were often reliably captured across

samples. The algorithm takes longer on average due

to the additional step of separation of long and short

edges.

Table 2 shows that for S

, average EER for left

and right palms is 2.63% for across session compar-

isons and this markedly improves on the published

benchmark of 3.86% (Kabacinski and Kowalski,

2011) on the same database. BGM performs on par

with the published result 1.11% for within session

comparisons using S

. Note the large difference

between EERs of across session and within session

data. Genuine samples of palm veins captured from

different sessions tend to have large displacement

in captured area. This reduces the common overlap

area between genuine samples and hence causes a

lower score. Comparing our results to the published

benchmark illustrates that BGM performs better than

the benchmark when palm vein samples have only a

partial overlap area. Five out of the 10 features give

competitive EERs of under 5% for within session

comparisons while three of these features (S

, S

c1c2

) also give competitive EERs for across session

comparisons.

Table 4 shows that there are three topological fea-

tures, that when combined with S

give a compet-

itive matching performance with total misclassiﬁca-

tion rate under 1.9%. To understand how much better

feature combination does over using the single mea-

sure S

, we adopt the following procedure. From the

ROC curve when using single feature S

(see Fig-

ure 5) we locate the FNMR when using S

, at a FMR

value, close to that obtained using two features. If the

FNMR value using two features is actually lower than

Figure 5: The Receiver Operating Characteristic (ROC)

curve, showing the plot of False Match Rates (FMRs) ver-

sus False Non Match Rates (FNMRs) over a range of op-

erating thresholds. The matching algorithm uses a single

feature S

the corresponding one feature value, that particular

pair of features gives a better result than using a single

feature for comparison. The performance comparison

between two features and one feature S

is shown in

Table 3.

Table 3 shows that the topological feature pairs

, S

) , (S

, S

) and (S

, S

c1c2

) do better than us-

ing S

alone, with (S

, S

) showing the maximum gain

in performance.

7 CONCLUSION

This paper proposes a graph-based approach to repre-

sent a palm vein template and proposes the Biometric

Graph Matching algorithm (BGM) to match the tem-

plates. The method is tested on the PUT public palm

vein database where each palm vein has 12 samples

captures across 3 sessions. This gave us a good num-

ber of genuine comparisons that were within the same

ThePalmVeinGraph-FeatureExtractionandMatching

301

Table 3: Comparison of matching performance when using a pair of features and a single feature S

. The rows where the

FNMRs are in boldface show the feature pairs whose performance is better than using S

alone.

Left Palm Right Palm

Features FMR FNMR TE FMR FNMR TE

, S

0.37(±.07) 4.72(±0.38) 2.34(±0.17) 0.38(±0.06) 4.29(±0.25) 1.93(± 0.08)

, S

0.52(±0.09) 4.02(±0.35) 2.1(±0.13) 0.36(±0.08) 4.23(±0.24) 1.9(±0.07)

, S

0.38(±0.09) 3.90(±0.32) 1.97(±0.11) 0.29(±0.04) 3.93(± 0.22) 1.74(±0.07)

, S

0.41(±0.06) 4.23(±0.3) 2.12(±0.12) 0.44(±0.07) 4.26(±0.21) 1.96(±0.08)

, S

0.31(±0.08) 4.37(±0.35) 2.15(±0.12) 0.40(±0.08) 4.20(± 0.25) 1.92(±0.08)

, S

0.38(±0.1) 3.83(±0.29) 1.94(±0.09) 0.27(±0.03) 3.79(±0.22) 1.67(±0.08)

, S

c1c2

0.45(±0.08) 3.82(±0.3) 1.97(±0.12) 0.36(±0.03) 3.79(±0.23) 1.72(±0.09)

, S

0.51(±0.11) 4.02(±0.33) 2.1(±0.12) 0.46(±0.09) 4.15(±0.23) 1.93(±0.06)

, S

0.48(±0.1) 3.87(±0.35) 2.02(±0.13) 0.44(±0.07) 4.03(±0.22) 1.87(±0.08)

Table 4: Matching performance when combining S

with each of the other 9 features. The table represents the mean over

10 folds with the standard errors in brackets. All results are expressed as percentages. The rows corresponding to the

combinations with lowest TE are highlighted.

Feature Pairs FMR FNMR FMR (S

) FNMR (S

)

, S

Left 0.37 4.72 0.37 4.28

Right 0.38 4.29 0.38 4.06

, S

Left 0.52 4.02 0.51 3.98

Right 0.36 4.23 0.36 4.21

, S

Left 0.38 3.90 0.38 4.22

Right 0.29 3.93 0.23 4.46

, S

Left 0.41 4.23 0.41 4.16

Right 0.44 4.26 0.44 4.03

, S

Left 0.31 4.37 0.31 4.39

Right 0.40 4.20 0.4 4.03

, S

Left 0.38 3.83 0.38 4.22

Right 0.26 3.79 0.26 4.63

, S

c1c2

Left 0.45 3.82 0.45 4.13

Right 0.36 3.79 0.36 4.21

, S

Left 0.51 4.02 0.51 3.98

Right 0.46 4.15 0.46 3.96

, S

Left 0.48 3.86 0.48 4.04

Right 0.44 4.03 0.44 4.03

session and across different sessions, to estimate the

difference in performance when biometrics are cap-

tured in different sessions. Note that multi session

comparison is typical in a real-world scenario. The

proposed method beats the existing state of art results

on the PUT public palm vein database (Kabacinski

and Kowalski, 2011) for genuine samples captured

across sessions. This establishes BGM as an effec-

tive algorithm when genuine palm vein samples have

partial overlap.

We remark that as the graph edit algorithm we

used gave signiﬁcant results we did not need to try

other cost functions. However, for future work these

could be applied to help evaluate the efﬁciency trade-

off of graph matching.

The paper shows that using the graph represen-

tation and BGM matching approach for palm vein

graphs is practical and efﬁcient. The graph-based

measure based on edge count outperforms the mea-

sure based on vertex count. This reinforces the ad-

vantage of using the edge information in representing

the palm vein graph. We further demonstrate that by

combining graph features, the matching performance

can be improved over using just a single feature. The

ability to explore and combine different graph fea-

tures based on the spatial characteristics of the palm

vein graph is one of the biggest advantages of using

the graph-based approach.

ICISSP2015-1stInternationalConferenceonInformationSystemsSecurityandPrivacy

302

ACKNOWLEDGEMENTS

We thank the anonymous referees for comments

which improvedthe clarity of the paper. This research

was funded by ARC grant DP120101188.

REFERENCES

Chen, H., Lu, G., and Wang, R. (2009). A new palm vein

matching method based on icp algorithm. In Proceed-

ings of the 2nd International Conference on Interac-

tion Sciences: Information Technology, Culture and

Human, pages 1207–1211, New York, USA. ACM.

http://doi.acm.org/10.1145/1655925.1656145.

Dubuisson, M. P. and Jain, A. (1994). A modiﬁed haus-

dorff distance for object matching. In Proceedings

of the 12th IAPR International Conference on Pattern

Recognition, pages 566–568. IEEE.

Gaikwad, D. P. and Narote, S. P. (2013). Multi-modal bio-

metric system using palm print and palm vein features.

In Annual IEEE India Conference (INDICON), pages

1–5.

Horadam, K. J., Davis, S. A., Arakala, A., and Jeffers, J.

(2011). Fingerprints as spatial graphs: Nodes and

edges. In Proc. of International Conference on Dig-

ital Image Computing Techniques and Applications

(DICTA), pages 400–405, Noosa, Australia.

Kabacinski, R. and Kowalski, M. (2010). Human vein pat-

tern segmentation from low quality images - a com-

parison of methods. Image Processing and Communi-

cations Challenges 2, 84:105–112.

Kabacinski, R. and Kowalski, M. (2011). Vein pattern

database and benchmark results. Electronics Letters,

47(20):1127–1128.

Kumar, A. and Prathyusha, K. V. (2009). Personal au-

thentication using hand vein triangulation and knuckle

shape. IEEE Transactions on Image Processing,

9:2127–2136.

Lajevardi, S., Arakala, A., Davis, S., and Horadam, K.

(2013). Retina veriﬁcation system based on biometric

graph matching. IEEE Transactions on Image Pro-

cessing, 22(9):3625–3635.

Lajevardi, S., Arakala, A., Davis, S., and Horadam, K.

(2014). Hand vein authentication using biometric

graph matching. IET Biometrics. doi: 10.1049/iet-

bmt.2013.0086.

Perona, P. and Malik, J. (1990). Scale-space and edge detec-

tion using anisotropic diffusion. IEEE Transactios on

Pattern Analysis and Machine Intelligence, 12:629–

639.

Riesen, K. and Bunke, H. (2009). Approximate graph

edit distance computation by means of bipartite graph

matching. Image and Vision Computing, 27(7):950–

959.

Shahin, M., Badawi, A., and Kamel, M. (2007). Biometric

authentication using fast correlation of near infrared in

hand vein patterns. International Journal of Biomedi-

cal Sciences, 2:141–148.

Wang, L., Leedham, G., and Cho, S. Y. (2007). Infrared

imaging of hand vein patterns for biometric purposes.

IET Computer Vision, 1(3-4):113122.

Watanabe, M., Endoh, T., Shiohara, M., and Sasaki, S.

(2005). Palm vein authentication technology and its

applications. In Proc. of Biometrics Consortium Con-

ference, pages 1–2, Arlington, VA.

Wenxiong, K. and Qiuxia, W. (2014). Contactless palm

vein recognition using a mutual foreground-based lo-

cal binary pattern. IEEE Transactions on Information

Forensics and Security, 9:1974–1985.

Zuiderveld, K. (1994). Contrast limited adaptive histogram

equalization. Academic Press Professional, Inc., San

Diego, CA, USA.

ThePalmVeinGraph-FeatureExtractionandMatching

303