Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

Introduction

In general, a palmprint recognition system consists of

three components: (1) preprocessing, (2) feature extrac-

tion, and (3) matching. A palmprint image will be ﬁrst

preprocessed to determine the region of interest. This

process includes segmentation and normalization so

that a canonical palmprint image can be produced for

sequel processing. Then, an intermediate palmprint rep-

resentation has to be derived through a feature extrac-

tion stage. The common palmprint representation

techniques can be mainly divided into ﬁve categories

according to the palmprint features, such as (1) line-

based approaches, (2) texture-based, (3) appearance-

based approaches, (4) multiples feature approaches,

and (5) orientations-based approaches [1]. A decision

is sought eventually through the matching process.

The matching is a task to calculate the degree of simi-

larity of two given palmprint and return a decision.

The establishment of palmprint matching strategy

is not a trivial problem. The primary reason for study-

ing of palmprint matching is due to the large deviation

in different impressions of the same palmprint, i.e.,

large intra-class variations. The possible factors that

contributed to intra-class variations are as follows:

1. Noise. It is introduced by the palmprint acquisition

system, e.g., residual prints left on palmprint acqui-

sition device

2. Adverse environmental condition such as bad illu-

mination condition

3. Incorrect interaction by the user with the palm-

print recognition system

4. Preprocessing and feature extraction errors. The pre-

processing and feature extraction algorithm are

imperfect and often introduce measurement errors.

The errors will be propagated from the prepro-

cessing stage (e.g., incorrect segmentation and

normalization of region of interest) to feature

extraction. Another instance is that in low-quality

palmprint images, the line-based feature extraction

process may introduce a large number of spurious

lines or miss to detect the true palm lines

On the other hand, palmprint images from different

palms may appear quite similar, i.e., small inter-class

variations, especially in palm print principle lines.

A well designed matcher attempts to ﬁnd the ‘‘best

ﬁt’’ between the two palmprint representations, and

thus reduces the errors that are introduced by the

above sources.

Furthermore, palmprint matching and feature

extraction are usually related to each other for both

veriﬁcation and identiﬁcation problems. Identiﬁcation

problem treats the searching for an input palmprint in

a database of M palmprints, thus it can be implemen-

ted as a series of consecutive execution of M one

to one matches (veriﬁcation) pairs of palmprints.

Several automatic palmprint matching algorithms

have been proposed in the literature of biometrics.

Based on the aforementioned feature representations of

palmprint image, the palmprint matching approaches

can be broadly classiﬁed into two categories:



Geometry-based matching: It is a natural way to

represent the feature of palmprint using geometry

objects, such as points, orientation, lines etc. Lines

in palmprint such as principle lines (e.g., head line,

life line, and hearth line) and coarse wrinkle are

the basic feature of palmprint [2]. A set of feature

points or lines segments along the basic palm lines

and/or the associated orientations can be extracted

from a palmprint image. Geometry-based match-

ing consists of ﬁnding the geometrical alignment

between the template and the input feature set that

returns in the maximum number of features pair-

ing or smallest/biggest degree of

▶ similarity/dis-

similarit y. However, line features is more popular

and widely used as it is relatively easier to be

extracted compared to point features in palmprint,

even in low spatial resolution palmprint image.



Feature-based matching: Geometry-based matching

approach relies heavily on the extraction of points

or lines feature from a palmprint image, this might

be difﬁcult in very low-quality palmprint images.

Alternatively, other features of the palmprint image

such as magnitude and orientation information in

palm lines can be utilized. The magnitude and

orientation of palm lines or texture, in general, can

be modeled and extracted by using appearance-

based [3, 4], transform-based [5], texture-based

[6, 7], and orientation-based techniques [8 , 9].

The approaches belonging to this category compare

palmprint features based on the similarity/dissimi-

larity measurement, such as Euclidean distance,

hamming distance, angular distance etc. between

the two corresponding feature vectors/matrices.

1050

Palmprint Matching

Besides the above two major categories, some other

techniques have also been proposed in the liter ature.

For instance,

▶ machine-learning based recognition

techniques such as Neural Networks [4], Support

Vector Machine [10], and Correlation-based matching

[11]. In principle, they could be regarded as sub-

categories of feature-based matching according to the

feature used, but it is more appropriate to categorize

them separately based on the matching approach.

Geometry-Based Matching

Point-Based

In view of structural similarity of ﬁngerprint and palm-

print impressions, where both are composed by ridges,

it is straightforward to adopt ﬁngerprint minutiae moti-

vated point matching approach to palmprint image.

A representative example of point-based matching was

proposed by Duta et al. [12]. In their method, given a

palmprint image, an average ﬁlter is ﬁrst used to

smooth the image followed by binarization process

based on a chosen threshold value, t. All pixels whose

value greater than t are labeled as palm line pixels while

others are regarded as background pixels. A set of

consecutive morphological erosions, dilations, and

subtractions are performed to eliminate the spu rious

palm lines. The outstanding foregrounds pixel loca-

tions are subsampled to retrieve a set of 200–400

pixel locations that will be considered a s feature points.

For each feature point, the orientation of its

corresponding palm line is calculated. The two sets of

feature points/orientations are geometrically aligned.

The goal of alignment of two feature point sets, A and

B is that to determine a corresponding feature points

in A such that the highest degree of correlation can be

found between A and B with respect to an optimal

transformation, T. The matching is ﬁnally carried out

based on a matching score that is deﬁned as a tuple

(P, D), where P is the percentage of point correspon-

dences with respect to the minimum number of feature

points in set A and B, and D is the average distance

between the corresponding points. This matching score

was devised based on two sources of intra-class varia-

tions, i.e., noise introduced by feature point extraction

and non-linear palm deformation due to various

ﬁnger positions [12], which are modeled by P and D,

respectively. It was shown that in a small size inked

palmprint database with 30 subjects, 95% of recogni-

tion rate was attained.

Another work on feature point matching was

reported by Jane et al. [13]. This work applied an

interesting point detector, namely Plessey operator to

extract the feature points. The interesting point detector

differs from the conventional edge detector in the sense

that the points detection is based on the how interes-

ting a point is. ‘‘Interesting’’ here refers to a set of

application-speciﬁc speciﬁcations that enables the oper-

ator to extract only those representative and distinctive

feature points for matching purpose. In general, inter-

esting point detectors will be operated based on a

three-step procedure. The ﬁrst step is to determine

a pre-speciﬁed size window, based on the average gradi-

ent magnitude. This is followed by the classiﬁcation

that distinguishes the types of singular points such

as corners, rings, spirals etc. based on a statistical test.

The last step is to reﬁne the located point within

the window.

For matching purpose, Hausdorff distance is

adopted to calculate the degree of similarity of

two feature points sets. The Hausdorff distance is

a non-linear operator, which measures the degree

of the mismatch between two feature point sets A

and B. Mathematically, the Hausdorff distance is

¼ max {h(A, B), h(B, A)} where h(A,

B) ¼ max

a2A

fmin

b2B

fdða

; b

Þgg and h(A, B) 6¼ h(B, A).

and b

are feature points of set A and B, respectively

and d(a

, b

) is an arbitrary metric between these

points. The major advantage of using the Hausdorff

distance for matching is that the computation can be

accelerated by partitioning the feature point sets A into

several subsets and match the B on these subsets simul-

taneously. To be speciﬁc, the feature point sets can be

represented in binar y matrices, A(i, j) and B(i, j). The

(i, j)th entry indicates an interesting feature point

position, which is set to 1 and 0 otherwise.

Two distance matrices, D

and D

are deﬁned as the

distance of each (i, j) location entry to the nearest

non-zero location entry of A and B, respectively.

Therefore, the Hausdorff distance, as a function of

translation can be computed by considering the

point-wise maximum of all the translated D

and D

in the form of d

(i, j) = max (a, b) where a ¼ max

– i, a

– j) and b ¼ max

þ i, b

þ j)[13].

Palmprint Matching

1051

A limitation of the feature points matching techni-

ques, which based on exhaustive scanning approach

is cumbersome and may not meet the real-time

requirement for on-line matching in a large database.

Line-Based

Compared to the point-based matching, line-based

matching is conceived more informative than the

point-based matching in palmprint recognition system

due to the rich line features in a palmprint image. Line-

base matching technique ﬁrst extract line feature,

which is composed by curves and straight line. In [2],

edge ﬁlters are applied to extract principle lines, thick

wrinkles, and ridges at various orientations repeatedly

and combine them with a post-processing algorithm

by line linking and thinning at the ﬁnal stage. The

representation of each extracted line segment is deter-

mined by a series of end points: (u

(i), v

(i)) and

(i), v

(i)), i =1,..., m where m is the number of

line segments. The end points coordinate is described by

a two-dimensional right coordinates system, which is

uniquely determined by the datum points. In general,

each line segment can be categorized by three para-

meters: (1) slope, m, (2) intercept, c and (3) angle of

inclination, a. They can be calculated using equations

m(i)=(v

(i)  v

(i))/(v

(i)  v

(i)), c(i)=v

(i) 

(i)m(i), a(i) = tan

1

(m(i)). The Euclidean distances

between the endpoints of two line segments, d

and d

given as d

ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ

ðu

ðiÞu

ðjÞÞ

þðv

ðiÞv

ðjÞÞ

and

ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ

ðu

ðiÞu

ðjÞÞ

þðv

ðiÞv

ðjÞÞ

For matching purpose, several conditions were

proposed as follows: (1) If both d

and d

are less

than a predeﬁned threshold value, t, then this implies

that both line segments are the same. (2) If the differ-

ence between a and c is less than sum of their respective

threshold values, then this implies two line segments

have the equal a and c. Among the class of equal a

and c, if one of d

and d

is less than t, then two

line segments are considered identical. (3) When

two line segments overlap, they are regarded as a single

line segment if the midpoint of one line segme nt is

between two endpoints of another line. Based on the

three rules given, a corresponding pair of palm lines

can be obtained. A decision criterion is hence deﬁned

as r ¼ 2N/(N

+ N

), 0 < r < 1 where N is the

number of these corresponding pairs; N

and N

are

the numbers of the line segments determined from two

palmprint images, respectively. In a medium scale

inked palmprint database that consists of 200 subjects,

it was reported that 92% of recognition rate can be

achieved.

Geometry-based matching approach, esp ecially

line features is an active research area and still evolving.

The researchers believe that line-based features in

palmprint are highly discriminant. Huang et al. ev i-

dences that even simple line-based features such as

principle lines also can show the high discriminability

[1]. In this work, they proposed to use a modiﬁed

ﬁnite Radon transform to extract the principle lines

and represent them in a binary matrix with size h  k,

where principle line point is set to 1 and 0 for others.

A new matching strategy, known as pixel-to-area com-

parison was devised bas ed on the pixel to area compar-

isons for robust line matching . Given two principle

line matrices, A and B, the matching score from A to

B is deﬁned as

S ¼ maxðsðA; BÞ; sðB; AÞÞ; 0  S  1

where

sðA; BÞ¼

i¼1

j¼1

Ai; jðÞ\Bi; jðÞ

and

sðB; AÞ¼

i¼1

j¼1

Bi; jðÞ\Ai; jðÞ

Here, \ denotes a logical ‘‘AND’’ operation, N

and

are the number of points on detected principle lines

in A and B.

Bi; jðÞis a small area around B(i, j ) and

is deﬁned as B(i +1,j), B(i1, j), B(i, j), B(i, j +1)

and B(i, j1). The same deﬁnition applies to

Ai; jðÞ.

S is devised in such a way that it is robust to slight

translations and rotations between the two images,

with limited to one pixel translation and 3



rotation.

In practice, the translation might be large due to im-

perfect preprocessing. This problem can be alleviated

by translating one image vertically and horizontally

repeatedly in the range of 2 to 2 pixels and match

with another image. The maximum value of S is

regarded as a ﬁnal decision score. The experiment

result showed that the method could achieve an equal

error rate (EER) of 0.565% in a large scale database

1052

Palmprint Matching

with 386 palmprints. It is believed that the perfor-

mance can be improved further if other line features

such as wrinkles are included.

Feature-Based Matching

Feature-based matching utilizes magnitude and orien-

tation information of palm lines, or texture in general

for matching purpose. Palm lines magnitude informa-

tion can be modele d and extracted by using statistical

and algebraic techniques such as appearance-based

feature representation [3, 4] (e.g., principle com-

ponent analysis (PCA ), ﬁsher discriminant analysis

(FDA), Independent component analysis (ICA) etc),

Fourier spectrum [5], wavelet transform [ 10], discrete

cosine transform [14], and convolution masks [13].

On the other hand, orientation information of palm

lines can be effectively extracted by using Gabor ﬁlters

[6, 8] and ordinal representation [9].

The common feature representation based on

statistical and algebraic techniques usually appear in

either one-dimensional feature vector with length n,

v ={v

| i =1,..., n} or two-dimensional feature matrix

with size m  n, V ={V

| i =1,..., m, i =1,..., n}.

For instance, [14] characterize a palmprint image by

using a set of context based wavelet signatures. Speciﬁ-

cally, a palmprint image is decomposed into J wavelet

scales and only three detail wavelet sub-band coefﬁ-

cients, i.e., horizontal, vertical, and diagonal. For each

sub-band coefﬁcient, four statistical readings, namely

(1) the Average Gravity Center Signature (AGCS), (2)

the Density Signature (DS), (3) the Spatial Dispersivity

Signature (SDS), and (4) the Energy Signature (ES) can

be computed. Each palmprint image will generate a

feature vector with length k = 3 + 3*3*J, which consists

of three pairs of AGCS, 3J DS, 3J SDS and 3J ES.

On the other hand, algebraic techniques such as

PCA, FDA, ICA etc ﬁrst transform the palmprint train-

ing images into a small set of characteristic feature

images, called Projection Matrix, which are the eigen -

vectors of the training set. Then, feature extraction

is performed by projecting a new palmprint image

with length n into the subspace spanned by the Pro-

jection Matrix. The output is a feature vector with

length k << n.

For the representation in one-dimensional feature

vector, the matching is normally done by using

distance metric such as city block distance, d

=|v

–

|, Euclidean distance, d

ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ

 v



, weighted

Euclidean distance, d

ﬃﬃﬃﬃﬃﬃ

 v



, angular

distance, d

etc., where i 6¼ j and w is a weight

vector. Without loss of generality, the degree of simi-

larity/dissimilarity of these distance metrics is given in

term of score between 0 and 1.

Another popular palmprint feature extractor used in

extracting the texture information is 2-D Gabor ﬁlter

and its variants [6–8]. In this technique, a 2-D Gabor

ﬁlter will convolute with a preprocessed palmprint

image followed by a robust encoding to convert the

con v oluted output into two binary matrices, which

corresponding to real and imaginary parts of the output.

Given two palmprint feature binary matrices, A

)

and B

) with size h  h, the matching is performed

via a normalized hamming distance H

such as

ðA

ði; jÞB

ði; jÞþA

ði; jÞB

ði; jÞÞ

where  is an bitwise Ex-OR operator [6 ]. The equa-

tion can be modiﬁed to cattle the translation prob-

lem with

¼ min

jsj<S;jtj<T



minðh;hþsÞ

i¼maxð1;1þsÞ

minðh;hþtÞ

i¼maxð1;1þtÞ

ði þ s; j þ tÞB

ði; jÞ

þ A

ði þ s; j þ tÞB

ði; jÞ



2HðsÞHðt Þ



s and t are set to 2 based on the assumption that

translation (both vertically and horizontally) is limited

in the range of 2 pixels and H(s) = min (h, h + s) – max

(1, 1 + s). However, this metric does not consider

rotation invariant, but this issue can be alleviated

during the enrolment stage. For instant, rotate the

coordinate system by a few degrees and perform fea-

ture extraction [6].

If a palmprint is not segmented properly during the

preprocessing stage, a number of non-palmprint pixels

will be introduced in the extracted feature matrix. In

this circumstance, these pixels are detected by using

some simple thresholding methods and their location s

can be recorded in the mask matrices, A

and B

that

corresponded to the feature matrices A and B. H

can

Palmprint Matching

1053

be modiﬁed to

\ B

\ A

 B

ðÞþA

\ B

\ A

 B

ðÞ

\ B

½7:

On the other hand, [8] perceived that orientation

information from the palm lines could be the promi-

nent features for a palmprint image. An even Gabor

ﬁlter with six different orientations is used to convo-

lute with the palmprint image and their contrast

magnitudes are sought. Based on the winner-take-all

competitive principle, the index (ranging from 0 to 5)

of the minimum contrast magnitude is represented by

three bits, namely competitive code. The matching of

two inputs, A and B can be carried out through

XXX

k¼0

\ B

\ A

 B



\ B

;

where A



is the ith bit plane of A(B).

As far as the performance is concerned, orientation-

based feature representation couples with matching

in hamming domain are deemed to be the most

promising techniques in palmprint recognition. In a

comparison study done in PolyU (medium size) and

UST datasets (large size), the competitive code showed

better performance with EER = 0% and EER = 0.38%,

respectively compare to PalmCode (0.34%, 1.68%)

and FusionCode (0.11%, 0.75%). However, the Ordi-

nal code, which is another orien tation-based feature

representation combined with robust encoding show

the best performance in terms of EER ¼ 0% and

EER ¼0.22% [9]. In addition to that, low computation

complexity and small template representation are also

another compelling advantage of this approach.

Summary

The estab lishment of palmprint matching methodolo-

gy is highly related to feature representation of palm-

print image and it is essential as the preprocessing

and feature extraction are imperfect. Geometry-based

matching approach was ﬁrst proposed in view of

natural representation of palmprint in feature points

and lines basis. However, it was believed that they are

less reliable as features point and lines are difﬁcult to be

explicitly extr acted, but recently it is receiving renewed

interest. Feature-based matching approach, as an

alternative, which utilizes the texture information

(magnitude and orientation of palm lines) of palm-

print, has also shown the signiﬁcan t advantages in

terms of performance, representation compactness,

and low computation complexity. The integration of

two approaches, e.g., in hierarchical manner [15] could

be the promising way to improve the palmprint recog-

nition systems in performance and speed.

Related Entries

▶ Authentication

▶ Identiﬁcation

▶ Veriﬁcation

References

1. Huang, D., Jia, W., Zhang, D.: Palmprint veriﬁcation based on

principal lines. Pattern Recognit. 41(4), 1316–1328 (2008)

2. Zhang, D., Shu, W.: Two novel characteristics in palmprint

veriﬁcation: datum point invariance and line feature matching.

Pattern Recognit. 32(4), 691–702 (1999)

3. Tee, C., Teoh, A.B.J., Goh, M.K.O., Ngo, D.C.L.: An automated

palmprint recognition system. Image Vis. Comput. 23(5),

501–515 (2005)

4. Shang, L., Huang, D.S., Du, J.S., Zheng, C.N.: Palmprint recog-

nition using FastICA algorithm and radial basis probabilistic

neural network. Neurocomputing 69, 13–15 (2006)

5. Li, W.X., Zhang, D., Xu, Z.Q.: Palmprint Identiﬁcation by

Fourier Transform. Int. J Pattern Recognit. Artif. Intell. 16(4),

417–432 (2002)

6. Kong, W., Zhang, D., Li, W.: Palmprint feature extraction

using 2-D Gabor ﬁlters. Pattern Recognit. 36(10), 2339–2347

(2003)

7. Zhang, D., Kong, W., You, J., Wong, M.: Online Palmprint

Identiﬁcation. IEEE Trans. Pattern Anal. Mach. Intell. 25(9),

1041–1050 (2003)

8. Kong, W., Zhang, D.: Competitive coding scheme for palmprint

veriﬁcation. In: Proceedings of the 17th International Confer-

ence on Pattern Recognition. 1, 520–523 (2004)

9. Sun, Z.N., Tan, T.N., Wang, Y.H., Li, S.Z.: Ordinal palmprint

representation for personal identiﬁcation. In: Proceedings of

IEEE International Conference on Computer Vision and Pattern

Recognition. 279–284 (2005)

1054

Palmprint Matching

10. Zhou, X.H., Peng, Y.H., Yang, M.: Palmprint Recognition Using

Wavelet and Support Vector Machines. Lect. Notes Comput. Sci.

4099, 385–393 (2006)

11. Hennings-Yeomans, P.H., Kumar, B.V.K.V., Savvides, M.: Palm-

print Classiﬁcation Using Multiple Advanced Correlation Filters

and Palm-Speciﬁc Segmentation. IEEE Trans Inf Forensic Secu-

rity 2(3), 613–622 (2007).

12. Duta, N., Jain, A.K., Mardia, K.V.: Matching of palmprints.

Pattern Recognit. Lett. 23(4), 477–485 (2002)

13. You, J., Li, W., Zhang, D.: Hierarchical palmprint identiﬁcation

via multiple feature extraction. Pattern Recognit. 35(4), 847–859

(2002)

14. Kumar, A., Zhang, D.: Personal recognition using hand-shape

and texture. IEEE Trans. Image Process. 15(8), 2454–2461 (2004)

15. You, J., Kong, W., Zhang, D., King Hong Cheung.: On hierar-

chical palmprint coding with multiple features for personal

identiﬁcation in large databases. IEEE Trans. Circ. Syst. Video

Tech. 14(2), 234–243 (2004)

Palmprint Recognition, 3D

▶ Palmprint, 3D

Palmprint Representation

▶ Palmprint Features

Palmprint Sensor

▶ Fingerprint, Palmprint, Handprint and Soleprint

Sensor

Parallel Fusion Network

A parallel fusion network is a fusion network topology

in which the sensor nodes are connected directly to the

central fusion processor. There is no distribu tion of the

processing through the network.

▶ Fusion, Decision-Level

Parametric Models

Synonym

Mathematical models

Definition

Parametric or mathematical models are compact math

and algorithmic representations that use variables that

when deﬁned allow the creation of the modeled phe-

nomenon that closely approximate the actual phenome-

non that is being modeled. An example of an empirically

derived parametric model is the University of Bologna’s

SFinGe ﬁngerprint generator.

▶ Biometric Sample Synthesis

Parametric-Based Biometrics

▶ Biometric Sample Synthesis

Part-Based Face Recognition

▶ Face Recognition, Component-Based

Partial Occlusion

Occlusions of a local region of the face with objects

such as sunglasses, scarf, hands, and hair are generally

Partial Occlusion

1055

called partial occlusions. Partial occlusions can, in

theor y, correspond to any occluding object. Generally,

the occlusion has to be of less than 50% of the face to

be considered a partial occlusion.

▶ Face Recognition, Component-Based

Passive Biometrics

In passive biometrics systems the subject does not have

to take active part in the process of identiﬁcation/

veriﬁcation or, in fact, does not even know that the

process of identiﬁcation takes place. In such biometrics

the user does not have to cooperate with the system

and does not need to touch any device or perform any

action.

▶ Ear Biometrics

Patron Format Specification

▶ Common Biometric Exchange Formats Framework

Standardization

Pattern Recognition

Pattern recognition aims to classify data (patterns)

based on either a prior knowledge or on statistical

information extracted from the patterns. The patterns

to be classiﬁed are usually groups ‘‘i’’ of measure-

ments or observations, deﬁning points in an appro-

priate ‘‘i’’ multidimensional space. A complete pattern

recognition system consists of a sensor that gathers the

observations to be classiﬁed or describe d; a feature

extraction mechanism that computes numeric or sym-

bolic information from the observations; and a classi-

ﬁcation or description scheme that does the actual job

of classifying or describing observations, relying on the

extracted features.

▶ Universal Background Models

PCA (Principal Component Analysis)

▶ Deformable Models

▶ Face Alignment

▶ Face, Forensic Evidence of

▶ Hand Shape

▶ Linear Dimension Reduction

▶ SFinGe

▶ Soft Biometrics

Pedestrian Detection

▶ Human Detection and Tracking

Peg

Short projecting pin used for marking position. Within

Hand Geometry biometrics, they may be placed on

a planar surface, to guide the user when placing his

or her hand.

▶ Hand Geometry

Pen Altitude

Angle measured counter-clockwise from the per-

pendicular projection of the pen onto the writing

plane to the pen. Altitude values can be acquired

1056

Passive Biometrics

as time-series data. Altitude is a factor representing the

pen inclination, and it is used together with azimuth.

These features are used for on-line signature

veriﬁcation.

▶ Signature Recognition

Pen Azimuth

Angle measured clockwise from the positive y axis to the

perpendicular projection of the pen onto the writing

plane. Azimuth values can be acquired as time-series

data. Azimuth is a factor representing the pen inclina-

tion, and it is used together with altitude. These features

are used for on-line signature veriﬁcation.

▶ Signature Recognition

Pen Inclination (Pen Tilt)

Angle characterizing how the pen is held.

How the pen is held depends on the per son, and it

changes during the signing process. Thus, this feature

is useful for on-line signature veriﬁcation. Pen inclina-

tion has two degrees of freedom and it can be repre-

sented in two ways. One way is to represent it by angles

measured from two axes. The other way is to represent

it by azimuth and altitude.

▶ Signature Recognition

Pen Pressure

Force of the pen on the writing surface.

Pen pressure values can be acquired as time-series

data. This feature is used for on-line signature

veriﬁcation. Generally, pen pressure is measured with a

stylus pen and tablet, or a pressure sensitive pen. Pen

pressure is represented by N quantized levels, for exam-

ple 1024 or 512. If the pen pressure is 0, the pen does not

touch the writing surface. Some kinds of tablet can

detect the pen position even when the pen does not

touch the tablet.

▶ Signature Recognition

Pen Tablet

A di gitizing or pen tablet is a ﬂat device that allows

recording handwriting movements. Usually these

devices are based on an electromagnetic principle.

The tablet has an embedded wire grid which acts as a

transmitter. The pen (which is speciﬁcally designed for

the tablet) acts as an antenna, which resonates and

emits a signal that is captured by the tablet, allowing

to detect its position with high accuracy. This allows

the tablet to detect the pen movement even if it is not

in contact with the tablet (in a reasonable range of

proximity).

▶ Signature Databases and Evaluation

Penetration Rate

The ratio of ﬁngerprints retrieved over the size of the

database.

▶ Fingerprint Indexing

Pen-tip Position (Pen Coordinates)

Position of the pen-tip on the writing plane.

The pen-tip position is generally represented using

Cartesian coordinates (x; y). The time series of

Pen-tip Position (Pen Coordinates)

1057

x-coordinates x(t) and y-coordinates y(t)areacquired

by a capture device at a sequence of sample points.

x(t) represents the position on the horizontal axis

and y(t) represents the position on the vertical axis

at each time t. This feature is used for on-line signa-

ture veriﬁcati on. Concatenating consecuti ve sample

points (x(t);y(t)) reproduces the shape of the

signature.

▶ Signature Recognition

Perceptual Expertise

Perceptual expertise is the change in perception due to

experience that may come about through a variety of

mechanisms such as shifting attention to different

dimensions or features, or new processes such as con-

ﬁgural processing.

▶ Latent Fingerprint Experts

Performance Bias in Synthesized

Biometric Data

Performance bias as applied to synthesized biome tric

data is an asystematic error experienced by a biometric

system. This type of bias can arise from a lack of or an

excessive presence of some synthetically generated dis-

tortions that appear in a larger amount or do not

present at all in real biometric data.

▶ Iris Sample Synthesis

Performance Evaluation Measures

▶ Performance Measures

Performance Evaluation,

Overview

SHIGUANG SHAN

,XILIN CHEN

,WEN GAO

1,2

Institute of Computing Technology, Chinese

Academy of Sciences, Beijing, China

Peking University, Beijing, China

Synonyms

Biometric technology test; Performance testing

Definition

Performance evaluation assesses accuracy and usabil-

ity of biometric algorithms or systems. Performance

measures are computed for veriﬁcation, identiﬁca-

tion, and watch list tasks, in order to either discover

the state-of-the-ar t of biometric technologies or

quantify how well a biometric system meets the

require ments of speciﬁc applic ations. Evaluation pro-

tocols and biometric databases for testing should

be c arefully designed to avoid biased results o r

conclusions.

Introduction

In the past several decades, many biometric algo-

rithms and commercial biometrics systems have

em erged. Many of them have reported very impres-

sive results on some public or private d atabases in

all kinds of publication s. However, it is hard to com-

pare them based onl y on the reported re sults, d ue

to the difference in either the evaluation methods or

the testing databases they exploited. Actually, the se

published results often lead to confusion in public:

application users are puzzled when choosing product

venders and researchers (especially green hands) may

not be able to clearly know the state of the art. There-

fore, it is indeed very important to standardize meth-

odology for performance evaluation of bio metric

technologies.

For most biometric technologies, there are two

main tasks when apply ing th em in practice, i.e.,

1058

Perceptual Expertise

veriﬁcation and identiﬁcation. The former needs to

answer ‘‘Is he who he says he is?’’, while the latter

cares a bout ‘‘who is he?’’. According to whe ther

the unidentiﬁed end-user is enrolled in the system,

identiﬁcation is further categorized into two t ypes:

▶ closed-set identiﬁcation and ▶ open-set identi-

ﬁcation. Typical veriﬁcation ap plication is access

con trol, while closed-set identiﬁcati on can be ap-

plied to mu g shot ret rieval for instance. In su rve il-

lance scenario, open-set identiﬁcation is also named

‘‘watch list,’’ which aims at answering ‘‘Is he one

of the persons of interest?’’ generally in real time.

For example, face recognition, gait recognition, and

speaker recognition can be applied for this purpose

in a surveillance scenario since they can work in non-

intrusive mode.

Performance evaluations can also be categorized

into three different types: algorithm evaluation, sce-

nario evaluation, and operational evaluation, as is

described in evaluation protocols part of this essay.

Performance Measures

Evidently, different tasks should explore distinct per-

formance measures. For veriﬁcation, receiver operating

characteristic (ROC) curve is generally used to show

the trade-off between two error rates: false reject rate

(FRR) versus false accept rate (FAR). Sometimes, the

equal error rate (EER) point on the ROC, where FRR is

equal to FAR, is used as a single measurement. As for

identiﬁcation, identiﬁcation rate, rank- k identiﬁcation

rate, or cumulative match characteristic (CMC) is

often used to compare different techniques. The reader

is referred to the

▶ performance measures entry for

more details.

For watch list applications, in some sense, it is

the veriﬁcation of the rank-1 identiﬁcation. So, its

performance can be measured by the identiﬁcation

rate at certain pre-deﬁned FAR, say 0.1%. This mea-

sure is exploited by face recognition vender test

(FRVT) [1].

Except the accuracy measures mentioned above,

there are also some performance criteria measur-

ing the usability of biometric systems, such as

▶ failure

to acquire rate,

▶ enrollment time, ▶ response time,

▶ throughput, and ▶ scalability. The readers are re-

ferred to the deﬁnitional entries for their description.

Datasets

The abovementioned performance measures are gen-

erally obtained by testing the biometric systems on

some databases. Evidently, the performance of a system

or algorithm depends not only on its capacity but also

the characteristic of the database. So, here it is worth

noting that pure recognition accuracy, say 100%,

means nothing if the database is not described clearly.

The following factors about validation database must

be considered carefully when performance evaluation

is conducted.

Here, ﬁrst several distinct datasets in evaluation:

training dataset, validation dataset, gallery, and probe

dataset need to be distinguished. Among them, the

training set is used for learning the biometric models

and designing the recognition algorithms, including

the feature extractor (e.g., principal component analy-

sis and discriminant analysis) and the classiﬁer. Vali-

dation set is used to tune the parameters of the leaned

models or the algorithms, for instance the dimension

of the feature vector or some empirical thresholds. In

some literature, training set and validation set are

combined together and called training set commonly.

For instance, in FVC2006 [2], a subset of ﬁnger-

print impressions acquired w ith various sensors was

provided to registered participants, to allow them to

adjust the parameters of their algorithms. The gallery

here means the dataset containing all the registered

biometric traits of all the enrolled users in the system,

that is, the templates for each enrolled users are

extracted from this dataset. Note that, the gallery

is often taken as part of (or the same as) the training

set by many researchers. This is mostly acceptable;

however, in many applications, each enrolled subject

may register only one biometric sample, which implies

that it is impossible to train a feature extractor (e.g.,

Fisher discriminant analysis) or classiﬁer. In this case,

a training set is necessary. The probe dataset con-

tains testing biometric samples that need to be recog-

nized by matching against the templates in the gallery.

Note that, for identiﬁcation task, all the subjects in

the probe set can be registered subjects (i.e., with at

least one template), while for veriﬁcation and watch

list applications , part of the subjects in the probe set

should be unregistered subjects, which is used as

impostors to estimate the false accept rate. In litera-

ture, the gallery and probe set are called together

Performance Evaluation, Overview

1059