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

Подождите немного. Документ загружается.

This method ﬁrst deﬁnes the average luminance of

each pixel as the average gray scale value of its sur-

rounding pixels. Then the facial images are converted

into binary images by comparing a pixel’s gray scale

value to its average luminance, with one or zero valu e

assigned by such pixel’s average luminance which is

above its gray value by a given margin, as illustrated by

the two feature maps in Fig. 1.

The skin area is then (usually below the eyes

and above the mouth)is div ided into several neighbor-

ing small blocks. For each block in the test image

the algorithm searches, in a local region of its cor-

responding block in the reference image, for the best

matching position. Figure 2 shows the matching results

of id entical twins, Ming and Gang , using Delean’s

Method. The upper row is one image of Ming and

two images of Gang. The lower row gives the pairwise

matching results and the corresponding similarity

scores. We use different colors to mark the continuous

regions after searching process. Same color indicates

that the neighbhoring blocks relative position change

is below a given threshold and they are considered as

continuous. More continuous block pairs implies that

the two images are likely to belong to a same person.

The probability can be formulated as a function of the

number of continuous block pairs. The underlying

assumption behind this method is: With two images

of a same person, the be st matching positions of the

neighboring blocks in the reference image tend to keep

the relative position of the blocks in the test image.

Otherwise, with two images of di fferent persons, the

relative position will not be kept. Finally the method

takes advantage of the continuity of block pairs and the

relative location of neighboring blocks to compute the

similarity score (Fig. 1).

The method is usually used as a complementary

to the other face matching algorithms rather than on

its own. Indentix reported that the incorporation of

the skin texture method into a local feature analysis

(LFA) matcher [3] could increase its accuracy by

20–25%. It was also reported [4] that this method

could differentiate between identical twins.

The Neven Vision method [5] uses Elastic Bunch

Graph [6] to represent faces and is capable of placing

this graph w ith hig h precision on a face in a presented

image. Based on the found facial landmarks, cor-

responding to facial landmarks like eye corner or the

tip of the nose, it ﬁrst accurately locates the areas

in the face that are used for skin texture analysis. The

skin areas are then warped and normalized before

matching. In the matching step, for each selected skin

patch, the feature vector is matched individually to the

face region of reference image and the most similar

skin patch is identiﬁed. The result is a more or less

distorted version of the graph in the original image.

From the similarity and distortion between the original

Skin Texture. Figure 1 Left: the blocks in a gallery skin texture image. Right: the best matching positions in a probe skin

texture image.

1230

Skin Texture

graph and the matched graph, a similarity score is

computed incorporating local geometric constraints

[5]. Having achieved an impressive performance in

FRVT2006 [7], Neven Vision also claims that their

skin texture analysis is about 1,000 times faster than

the Delean/Identix algorithm. This method works on

very high resolution images with inter-pupil distance

of at least about 600 pixels. When this condition is

fulﬁlled, it can achieve better recognition accuracy

than iris-based method and can even outperform

human performance [7].

Pros and Cons

Skin texture based methods have the following

advantages:



The facial skin has a ﬁne texture that is determined

randomly during embryonic gestation. Even iden-

tical monozygotic twins have completely indepen-

dent skin textures.



The acquisition of skin texture is similar to captur-

ing a photograph, and can be perf ormed from

several meters away. There is no need for the person

to be identiﬁed to touch any equipment that has

recently been touched by a stranger, thereby elim-

inating an objection that has been raised in some

cultures against ﬁnger-print scanners, where a

ﬁnger has to touch a surface, or retinal scanning,

where the eye can be brought very close to a lens.



Skin texture contains plentiful local detailed infor-

mation and can achieve a very high recognition

accuracy.

However, they have the following disadvantages:



High resolution is of vital importance for skin

texture based methods. The performance will dra-

matically drop with blurred images because the

image resolution is reduced.



Skin texture may not be the same w ith aging, exag-

gerative expression, pose variation, and illumina-

tion changes.



In surveillance environment, it is difﬁcult to cap-

tain high resolution face images at a far distance if

the person to be identiﬁed is not cooperative by

holding the head still and looking at the camera.

This may narrow possible applications of skin

texture methods.

Summary

As a new method for biometrics, skin texture has

demonstrated its effectiveness in identifying indivi-

duals in high resolution images. The increased demand

for accurate, reliable, fast, and convenient biometric

Skin Texture. Figure 2 An example of Delean’s method in matching of twins.

Skin Texture

1231

technologies, and availability of inexpensive comput-

ing resources and cameras will all beneﬁt to the rapid

technical advances in thi s direction. Although there

are several unsolved problems and disadvantages as

aforementioned, the skin texture method is surely

of signiﬁcance from both application and research

perspectives.

Related Entries

▶ Face Recognition

▶ Local Image Features

▶ Skin Spectroscopy

References

1. Cula, O., Dana, K., Murphy, F., Rao, B.: Skin Texture Modeling.

Int. J. Comput. Vis. 62(1–2), 97–119 (2005)

2. Delean, B., Escolls, R.: Method and apparatus for probabilistic

image analysis. u.s. patent application publication no. 2004/

0052418 (2004)

3. Penev, P., Atick, J.: Local feature analysis: a general statistical

theory for object representation. Network: Comput. Neural Syst.

7(3), 477–500 (1996)

4. Identix Turbo charges its Biometric Engineers, Biometric Tech-

nology Today, ISSN 0969-4765/04, 12(4), 1,12 (April 2004).

5. Adam, H., Neven, H., Steffens, J.: Single image based multi-

biometric system and method u.s. patent application publica-

tion no. 2006/0050933 (2006)

6. Wiskott, L., Fellous, J., Kruger, N., Malsburg, C.: Face recogni-

tion by elastic bunch graph matching. IEEE Trans. Pattern Anal.

Mach. Intell. 19(7), 775–779 (1997)

7. Phillips, P., Scruggs, W., O’Toole, A., Flynn, P., Bowyer, K.,

Schott, C., Sharpe, M.: Frvt 2006 and ice 2006 large-scale results

(2007)

Skull, Forensic Evidence of

MIN GQUA N ZHOU

Beijing Normal University, Beijing, China

Synonyms

Craniofacial reconstruction; Skull-photo superimposi-

tion; Craniofacial superimposition

Definition

Forensic evidence of skull is the technique of obtaining

the identi ty of a person based on attributes associated

with an individual skull. Generally, the methods of

determining the identity of a given skull can be classi-

ﬁed into two categories: craniofacial reconstruction,

which reconstructs the facial appearance from a given

skull based on the assumed relationship between the

soft tissue envelope and the underlying skull substrate

[1, 2] and skull-photo superimposition or craniofacial

superimposition, which superimposes the skull and

face photo in a certain way to verify whether they are

from the same person, by comparing their feature

points [3].

Introduction

In forensic pathology, it is esse ntial to determine

the identity of the dead. However, in many instances,

due to the storage environment and discovery time

of the corpse, the soft tissue becomes highly

decomposed, and forensics cannot determine the ap-

pearance of the dead directly. As a result, attempts

have been made to reconstruct the appearance from

the remaining skeletons, especially the skull. Research

has revealed that there are certain relationships be-

tween the facial soft tissue and the skull, and the skull

has evident inﬂuence on the shape, location and struc-

ture of the facial soft tissue. Thus, the facial appearance

is related to the skull to a certain extent, and it ensures

the ﬁdelity of the forensic evidence of skull [4, 5].

To determine the identity of a skull, an option is

to reconstruct the 3D facial appearance explicitly or

implicitly. The former is known as craniofacial recon-

struction, which reconstructs the 3D facial appearance

based on the given skull. Face is the most expressive

part of the human being. Reconstructing a realistic

3D facial appearance is a challenging task. The latter

method referred to as skull-photo superimposition

does not reconstruct the 3D facial appearance explicit-

ly. Instead, it superimposes the skull and photo at the

same imaging condition to check whether the given

skull and photo are of the same person, by matching

their feature points. As in the case of ﬁngerprint

and tooth-based inspection, skull-photo superi mposi-

tion-based inspection can also verify the identity, and

is an important method in

▶ forensic anthropology.

1232

Skull, Forensic Evidence of

In cases where the body is dismembered, skull-photo

superimposition becomes even more important and

unique in identity veriﬁcations. DNA can be of valu e

in these cases and even when only the bones remain.

Craniofacial Reconstruction

Traditionally, craniofacial reconstruction methods can

be classiﬁed into 2D reconstruction and 3D recon-

struction. In 2D reconstruction, the hand drawn

portrait of the dead is viewed by experienced anthro-

pologists, art ists or forensics based on the skull. 3D

reconstruction methods normally set so me marker

points with height on the skull model, and use clay as

a replacement of soft tissue, to shape the facial appear-

ance based on the skull according to anatomical

knowledge. However, these traditional reconstruction

methods have several limitations:

1. The who le process is very time-consuming, it

takes almost a month to reconstruct a person’s

appearance.

2. The result relies heavily on the expertise and expe-

rience of the reconstructors, the accuracy and reli-

ability of the reconstruction is affected by many

human factors.

3. The soft tissue and shape of the human face is

different for each human race and living environ-

ment, it is favorable to generate several recon-

structed facial appearances simultaneously when

the race of a given skull is not known. Using tradi-

tional reconstruction methods, however, only one

appearance for one race can be obtained.

4. It is difﬁcult to edit or modify the reconstructed

facial appearance once it is complete.

In the past decades, however, imaging technologies

▶ CT and MRI have been applied to convert the

skull into digital data so that it can be stored and

processed easily, leading to computer-aided craniofa-

cial reconstruction. In a typical computer-aided cra-

niofacial reconstruction system, the relationship

between the skull and soft tissue is not determined by

the expertise and experience of the reconstructors.

Instead, the relationship is learnt from the training

data, which are reference pairs of captured skull data

and soft tissue data. The knowledge of obtaining facial

detail from the skull and the distribution law of the soft

issues are obtained through feature point pairs on the

skull and face appear ance. This guides the creation of

an individual skull. Traditio nally, several methods are

available to obtain the training samples of reference

skull data and soft tissue data, for example, using

▶ acupuncture on the dead body to measure the thick-

ness of the soft tissue. Following advances in imaging

technologies, CT technologies may be used to collect

training data for craniofacial reconstruction more ac-

curately and efﬁciently. A pre-processing step, typically

consisting of ﬁltering and geometric transform a-

tion, is performed to eliminate the noise in each

2D CT slide and align the slides correctly [6, 7].

Sometimes, to facilitate 3D reconstruction of skull

and face, edge detection and edge tracking is per-

formed on the 2D CT slides to extract the contour

curves in each slide [8–10]. After pre-processing, 3D

reconstruction of the skull and face can be undertaken

to generate the training samples for the craniofacial

reconstruction system. In some systems, an explicit

model describing the relationship between the skull

and soft tissue is built from the training samples.

Along with the development of statistical learning

theor y, some systems instead reconstruct 3D facial

appearance directly from a given skull by using statis-

tical learning methods, without developing an explicit

model describing the knowledge of these training

samples. However, though craniofacial reconstruction

can be used as an auxiliary method to aid detectives

in their work, its result cannot be used as legal evi-

dence in a court of law.

Craniofacial reconstruction can also be used in

▶ palaeoanthropology, for example, to reconstruct

the facial appearance of famous anci ent people accord-

ing to their skull, in order to see how these people

looked and compare it with their portraits. Further,

Craniofacial reconstruction can be used in facial sur-

ger y to help simulate the process and effect of the

surgery, providing a more detailed and accurate sur-

ger y plan, and reducing the risk of operation.

Skull-photo Superimposition

Skull-photo superimposition was ﬁrst developed to

verify the ﬁdelity of portraits of ancient notables. In

the 1980s doubts were raised about the authenticity of

many of the portraits as it was believed that the pain-

ters wanted to please these notables. To verify this

conjecture, anatomists compared the skulls and

Skull, Forensic Evidence of

1233

portraits of notables. However, the earliest attempts to

use the skull for identity veriﬁcation in literature were

made at the end of the 19th century by Schau Ffhausen,

Von Froliep and other scientists. In 1935 Brash

recorded the ﬁrst successive use of craniofacial super-

imposition as forensic evidence in the Ruxton case.

Since then, along with the development of photo-

graphic and video technolo gies, many improvements

and upgrades have been made by researchers. In the

last 20 years, researchers have extended craniofacial

superimposition from photographic superimposition

to video superimposition, which has proven to be a

successful craniofacial identiﬁcation.

A typical skull-photo superimposition operation

can be separated into the following stages: First, the

skull data is obtained through a 3D scanner or recon-

structed from CT scans. After collecting the 3D skull

data, the difference between locations, sizes and orien-

tations of the skull and photo need to be minimized

for the superimposition task. As the photo is difﬁcult

to manipulate in three-dimension, the skull is nor-

mally edited or adjusted in three-dimension to be

aligned with the photo. Moreover, the facial photo

may need to exposed to several pre-processing steps

such image enhancement, scaling, and rotation [6, 7].

Next, facial features including head contour curves

are extracted from the facial photo [8–10]. In the

next stage marker points are selected, the traditional

method of manual selection relies on the expertise

of operators, and the result tends to be erroneous.

On the other hand, despite the recent advances in

computer vision, a fully automatic method of marker

points selection is not realistic at the current stage.

In practice, a hybrid system of combining manually

labeling and computer veriﬁcation yields a reason-

able result. Finally, skull-photo registration techniques

and standardized veriﬁcation process are performed

to verify the identit y [11–14]. Unlike craniofacial

reconstruction, skull-photo superimposition opera-

tions, if performed well, can be used as legal evidence

in court.

Related Entries

▶ Biometrics, Overview

▶ Forensic DNA Evidence

▶ Identiﬁcation

▶ Veriﬁcation

References

1. Prag, J., Neave, R.: Making Faces Using Forensic and Archaeo-

logical Evidence. British Museum Press, UK (1997)

2. Taylor, K.T.: Forensic Art and Illustration. CRC Press, New York

(2001)

3. Dorion, R.: Photographic superimposition. JFSCA. 28(3),

724–734 (1983)

4. Wilkinson, C.: Forensic Facial Reconstruction. Cambridge

University Press, Cambridge (2004)

5. Stephan, C., Henneberg, M.: Building faces from dry skulls: Are

they recognized above chance rates? J. Forensic. Sci. 46(3),

432–440 (2001)

6. Hummel, A.D.: Image enhancement by histogram transforma-

tion. Comput. Graph Image Process. 6(2), 184–195 (1977)

7. Ketcham, D.J.: Real-time image enhancement technique.

In: Proceeding SPIE/OSA, Conference on Image Processing,

pp. 120–125. Paciﬁc Grove, CA (1976)

8. Kass, M., Witbin, A., Tetzopoulos, D.: Snake: Active contour

model. Int. J. Comput. Vis. 1(4), 321–331 (1988)

9. Lai, K.F., Chin, R.T.: Deformable contours modeling and extrac-

tion. IEEE Trans. Pattern Anal. Mach. Intell. 17(11), 1084–1090

(1995)

10. Gunn, S.R., Nixon, M.S.: Snake head boundar y extraction

using global and local energy minimization. Proceeding of

international conference on pattern recognition (ICPR ’96),

pp. 581–585. Vienna, Austria (1996)

11. Besl, P., McKay, N.: A method for registration of 3-D shapes.

IEEE Trans. Pattern Anal. Mach. Intell. 14(2), 239–256 (1992)

12. Pelizarri, C.A., Chen, G., Spelbring, D.: Accurate three-dimen-

sional registration of CT, PET, and/or MRI images of the brain.

J. Comp. Assist. Tomogr. 13(1), 20–26 (1989)

13. Wells, W.M., Viola, P., Atsumi, H., Nakajima, S., Kikinis, R.:

Multi-modal volume registration by maximization of mutual

information. Med. Image Anal. 1(1), 35–51 (1996)

14. Vandermeulen, D., Collignon, A., Michiels, J., Bosmans, H.,

Suetens, P., Marchal, G.: Multi-modality image registration

within covira. Med. Image-Anal. 19, 29–42 (1995)

Skull-Photo Superimposition

▶ Skull, Forensic Evidence of

Slap Or Four-Four-Two device

It refers to a device used to capture the ten ﬁnge rprints

of a person using the following capture sequence:

1234

Skull-Photo Superimposition

simultaneous capture of the four ﬁngers (index, mid-

dle, ring, and little ﬁnger) of one hand, simultaneous

capture of the four ﬁnge rs of the other hand, and

simultaneous capture of the two thumbs. A segmenta-

tion algorithm is needed to detect and separate the

ﬁnger in ten single images.

▶ Fingerprint, Palmprint, Handprint and Soleprint

Sensor

Smart Cameras

Smart cameras are cameras that along with the sensor

have additi onal processing elements (micro-control-

lers, DSP) for processing data on-site. The results of

the processing are then transmitted to other nodes in

the network. Smart cameras form an important part of

the distributed computing idea, as only the relevant

information from each camera is transmitted to other

cameras/central processing agency. This alleviates the

burden on other nodes to replicate the processing as

well as the need for superior computing resources at a

central node.

▶ Surveillance

Smart Card

A smart card is a small plastic card with an embedded

microchip that can store and/or process information.

It can receive and submit data to or from any system

equipped with an appropriate card-reader module. In

commercial hand-geometry devices, for example,

smart cards are often used as storage media for user-

templates and as such eliminate the need for storing

templates in the internal memory of the device. Its

synonyms include integrated circuit card (IC C) and

chip card.

▶ Hand-Geometry Device

▶ Tamper-proof Operati ng System

Soft Biometrics

KARTHIK NANDAKUMAR

,ANIL K. JAIN

Institute for Infocomm Research, A*STAR,

Fusionopolis, Singapore

Michigan State University, East Lansing, MI, USA

Synonyms

‘‘Light’’ biometrics

Definition

Any anatomical or behavioral characteristic that pro-

vides some information about the identity of a person,

but does not provide sufﬁcient evidence to precisely

determine the identity can be referred to as a soft

biometric trait. Personal attributes like gender, ethnicity,

age, height, weight, eye color, scars, marks, tatoos, and

voice accent are examples of soft biometric traits. Soft

biometric information complements the identity infor-

mation provided by traditional (primary) biometric

identiﬁers such as ﬁngerprint, face, iris, and voice.

Hence, utilizing soft biometric traits can improve the

recognition accuracy of primary biometric systems.

Introduction

Systems that consolidate evidence from multiple sources

of biometric information (e.g., face, ﬁngerprint, hand

geometry, iris, etc.) in order to reliably determine the

identity of an individual are known as multibiometric

systems [1]. Multibiometric systems can alleviate many

of the limitations of unibiometric systems such as

nonuniversality and lack of distinctiveness, thereby

reducing the error rates signiﬁcantly. However, using

multiple biometric traits will increase the enrollment

and veriﬁcation times, cause more inconvenience to

the users, and increase the overall cost of the system.

An alternate way for reducing the error rates of the

biometric system without causing any additi onal in-

convenience to the user is to incorporate soft identi-

ﬁers of human identity like gender, ethnicity, height,

eye color, etc. into a (primary) biometric recognition

system [2]. Figure 1 depicts a scenario where both

primary (face) and soft (gender, ethnicity, height, and

Soft Biometrics

1235

eye color) biometric information can be automati-

cally extracted and utilized to verify an user’s identity.

In this scenario, the height of the user can be estimated

as he approaches the camera and his age, gender,

ethnicity, and eye color can be estimated from his

face image. These additional attributes can be used

along with the face biometric to accurately identify

the person.

Motivation for Utilizing Soft Biometric

Traits

The usefulness of soft biometric traits in improving

the performance of the primary biometric system

can be illustrated by the following example. Consider

three users A (1.8 m tall, male), B (1.7 m tall, female),

and C (1.6 m tall, male) who are enrolled in a ﬁnger-

print biometric system that works in the identiﬁca-

tion mode. Suppose user A presents his ﬁngerprint

sample X to the system, it is compared to the templates

of all the three users stored in the database and the

posteriori matching probabilities of all the three users

given the sample X are calculated. Assume that the

outputs of the ﬁngerprint matcher are P(AjX) ¼ 0.

42, P(BjX) ¼ 0.43, and P(CjX) ¼ 0.15. In this case,

user A will be falsely identiﬁed as user B based

on the Bayesian decision rule. However, assume that

as the user approaches the ﬁngerprint sensor, there

exists a secondary system that automatically identi -

ﬁes the gender of the user as male and measures the

user’s height as 1.78 m. This information, in addition

to the posteriori matching probabilities given by the

ﬁngerprint matcher, is likely to lead to a correct iden-

tiﬁcation of the user as user A.

The ﬁrst biometric system developed by Alphonse

Bertillon in 1883 used anthropometric features such as

the length and breadth of the head and the ear, length of

the middle ﬁnger and foot, height, etc. along with attri-

butes like eye color, scars, and tatoo marks for ascertain-

ing a person’s identity. These measurements were

obtained manually by Bertillon. Although each individ-

ual measurement in the Bertillonage system may exhibit

some (intra-user) variability, a combination of several

quantized (or binned) measurements was sufﬁcient to

manually identify a person with reasonable accuracy.

The Bertillon system was dropped in favor of the Henry’s

system of ﬁngerprint identiﬁcation over 100 years back

due to three main reasons: (1) lack of persistence – the

anthropometric features (e.g., height) can vary signiﬁ-

cantly for juveniles; (2) lack of distinctiveness – features

such as skin color or eye color cannot be used for distin-

guishing between individuals coming from a similar

ethnic background; and (3) the huge time, effort, and

training required to get reliable measurements.

While the anthropometric features used in the

Bertillon system provide some information about the

identity of the user, they are not sufﬁcient for accurately

identifying the user. Hence, these attributes can be

referred to as ‘‘soft biometric traits’’. Figure 2 shows

some examples of soft biometric traits. Since the soft

biometric information complements the identity infor-

mation provided by traditional (primary) biometric

identiﬁers such as ﬁngerprint, face, iris, and voice, utiliz-

ing soft biometric traits can improve the recognition

accuracy of primary biometric systems. Many practical

biometric systems collect soft biometric information

about the users during enrollment. For example, the

ﬁngerprint card used by the Federal Bureau of Investiga-

tion (FBI) includes information on the gender, ethnicity,

height, weight, eye color, and hair color of the person

along with the prints of all ten ﬁngers. However, the

searches in the FBI Automated Fingerprint Identiﬁcation

System (AFIS) are solely based on ﬁngerprints and the

potentially useful soft biometric information is either

ignored during the search or used only for manual

veriﬁcation after a short list of potential ﬁngerprint

matches is identiﬁed. If the soft biometric characteristics

can be automatically extracted and/or utilized during the

automated matching process, the overall accuracy of the

system will improve and the need for manual interven-

tion will be reduced.

Soft Biometrics. Figure 1 A scenario where the primary

biometric identifier (face) and the soft biometric attributes

(gender, ethnicity, eye color and height) are automatically

extracted and utilized to verify a person’s identity.

1236

Soft Biometrics

Challenges in Using Soft Biometrics

Two key challe nges need to be addressed to incorporate

the soft biometric information into the traditional

biometric framework. The ﬁrst challenge is the auto-

matic and reliable extraction of soft biometric informa-

tion in a nonintrusive manner, without causing any

inconvenience to the users. It must be noted that the

failure of Bertillon-like systems was caused by the

unreliability and inconvenience in the manual extrac-

tion of these features. Once the soft biometric infor-

mation about a user is available, the challenge is to

optimally combine this information with the

▶ primary

biometric identiﬁer so that the overall recognition ac-

curacy is enhanced. While soft biometric traits can be

used for

▶ ﬁltering a large database or for tuning the

parameters of a biometric system, such applications

require a highly accurate soft biometric feature extrac-

tion module. Since it is very difﬁcult to extract soft

biometric features with 100% accuracy, the information

fusion system needs to be designed in such a way that

the overall recognition accuracy is enhanced even when

the soft biometric feature extraction is not perfect.

Automatic Soft Biometric Feature

Extraction

Soft biometric traits are available and can be extracted in

a number of biometric applications. For example, attri-

butes like gender, ethnicity, age, and eye color can be

extracted with sufﬁcient reliability from the face images.

Automatic recognition of gender has been extensively

studied and a majority of the gender recognition sys-

tems proposed in the literature are based on frontal face

images. Furthermore, most of the these systems follow

an appearance-based approach to gender recognition

(see [3, 4] and the references therein). The face images

are typically cropped to include only the forehead, eyes,

nose, and mouth regions, and normalized for pose and

illumination changes. A pattern classiﬁer is directly

trained using the normalized face images to learn the

decision boundary between the male and female classes.

The accuracy of face-based gender recognition systems

is typically around 90% when presented with good

quality frontal face images. Some of the techniques

used for gender recognition can also be applied to

classify people based on their ethnicity [3, 5].

Automatic age determination is a more difﬁcult

problem than gender classiﬁcation. Kwon and Lobo

[6] presented an algorithm for age classiﬁcation

from facial images based on cranio-facial changes in

feature-position ratios and skin wrinkle analysis. More

recently, Lanitis et al. [7] performed a quantitative

evaluation of the performance of various classiﬁers

developed for the task of automatic age estimation

from face images. All the classiﬁers used eigenfaces

obtained using

▶ principal component analysis

(PCA) as the input features. Quadratic models, short-

est distance classiﬁer, neural network classiﬁers, and

hierarchical classiﬁer were used for estimating the age.

The best age estimation algo rithm had an average

absolute error of 3.82 years, which was comparable to

the error made by humans (3.64 years) in performing

the same task. Geng et al. [8] proposed an iterative

learning algorithm known as AGES for age estimation

from PCA features. The AGES algorithm achieved a

mean absolute error of 6.77 years in estimating the age

on a database with 1,002 face images obtained from

82 subjects.

Soft Biometrics. Figure 2 Examples of soft

biometric traits.

Soft Biometrics

1237

Gender, speech accent, and perceptual age of the

speaker can also be inferred from the speech signal.

The weig ht of an user can be measured by asking him

to stand on a weight sensor while he is providing his

primary biometric. Ailisto et al. [9] used body fat

measurement as a soft biometric trait. The height of

a person can be estimated from a real-time sequence

of images as the user approaches the biometric system.

Jain et al. [2] implemented a real-time vision system for

automatic extraction of gender, ethnicity, height, and

eye color. The system was designed to extract the

soft biometric attributes as the person approaches

the primary biometric system to present his primary

biometric identiﬁer (face and ﬁngerprint). Their soft

biometric system is equipped with two pan/tilt/zoom

cameras. Camera 1 monitors the scene for any human

presence based on the motion segmentation image.

Once camera 1 detects an approaching person, it mea-

sures the height of the person and then guides camera

2 to focus on the person’s face.

Fusion of Primary and Soft Biometric

Information

Jain et al. [2] developed a Bayesian framework for

fusion of primary and soft biometric features. The

main advantage of this framework is that it does

not require the soft biometric feature extractors to be

perfect (100% accurate). Assume that the primary

biometric system is based on R

 1) biometric

identiﬁers like ﬁnge rprint, face, iris, and hand geome-

try. Further, the soft biometric system is based on

 1) attributes like age, gender, ethnicity, eye

color, and height. Let o

, o

, ..., o

represent the M

users enrolled in the database. Let x ¼½x

; x

; ...; x



be the collection of primary biometric feature vectors.

Let p(x

) be the likelihood of observing the primary

biometric feature vector x

given the user is o

.If

the output of each individual modality in the pri-

mary biometric system is a set of match scores,

¼½s

1;k

; s

2;k

; ...; s

, one can appro ximate p(x

)

by p(s

), provided the genuine match score distri-

bution of each modality is known.

Let y ¼½y

; y

; ...; y

 be the soft biometric fea-

ture vector, where, for example, y

could be the gender,

could be the eye color, etc. The posteriori probability

of user o

given both x and y can be calculated by

applying the

▶ Bayes rule as follows:

Pðo

jx; yÞ¼

pðx; yjo

ÞPðo

pðx; yÞ

; ð1Þ

where P(o

) is the prior probability of observing user

. If all the users are equally likely to access the system

and if all the primary biometric feature vectors

; ...; x

and all the soft biometric variables

; y

; ...; y

are independent of each other given

the user’s identity o

, the discriminant function,

(x, y), for user o

, can be computed as

ðx; yÞ¼

j¼1

log pðx

Þþ

r¼1

log pðy

Þ: ð 2Þ

During the identiﬁcation phase, the input biomet-

ric sample is compared with the templates of all the M

users enrolled in the database and the discriminant

functions g

, ..., g

are computed. The test user is

identiﬁed as that user with the largest value of discrim-

inant function among all the enrolled users. A simple

method for computing the soft biometric likelihoods

p(y

),r ¼ 1, 2, ...,R

, k ¼ 1, 2, ..., M is to estimate

them based on the accuracy of the soft biometric

feature extractors. Jain et al. [2] also suggested the

use of a scaling factor b

,0 b

 1, to ﬂatten the

likelihood distribution of each soft biometric trait.

The scaling factor b

can act as a measure of the

reliability of the rth soft biometric feature and its

value can be set depending on the environment (hos-

tile or friendly) in which the system operates.

Summary

In addition to the match scores provided by the bio-

metric matchers, ancillary information may also be

available to a biometric system. Soft biometric char-

acteristics like gender, ethnicity, height, and weight

provide some information about the identity of the

user. Although the soft biometric information alone

is not sufﬁcient for accurate recognition, they can be

used to complement the information provided by the

primary bio metric identiﬁers like ﬁngerprint, iris, and

face. Techniques for automatically extracting soft bio-

metric information have been developed only recently.

Hence, fusion schemes that incorporate such ancillary

information have not been thoroughly explored and

there is a large scope for conducting more in-depth

research in this area.

1238

Soft Biometrics

Related Entries

▶ Biometrics, Overview

▶ Biometric Fusion

▶ Face Recognition

▶ Fingerprint Recognition

▶ Multibiometrics

References

1. Ross, A., Nandakumar, K., Jain, A.K.: Handbook of Multibio-

metrics. Springer, Berlin (2006)

2. Jain, A.K., Nandakumar, K., Lu, X., Park, U.: Integrating faces,

ﬁngerprints and soft biometric traits for user recognition.

In: Proceedings of ECCV International Workshop on Bio-

metric Authentication (BioAW), vol. LNCS 3087, pp. 259–269.

Springer, Prague, Czech Republic (2004)

3. Gutta, S., Huang, J.R.J., Phillips, P.J., Wechsler, H.: Mixture

of experts for classiﬁcation of gender, ethnic origin, and

pose of human faces. IEEE Trans. Neural Netw. 11(4), 948–960

(2000)

4. Moghaddam, B., Yang, M.H.: Learning gender with support

faces. IEEE Trans. Pattern Anal. Mach. Intell. 24(5), 707–711

(2002)

5. Lu, X., Jain, A.K.: Ethnicity identiﬁcation from face images. In:

Proceedings of SPIE Conference on Biometric Technology for

Human Identiﬁcation, vol. 5404, pp. 114–123. Orlando, USA

(2004)

6. Kwon, Y.H., Lobo, N.V.: Age classiﬁcation from facial images. In:

Proceedings of IEEE Conference on Computer Vision and

Patern Recognition, pp. 762–767 (1994)

7. Lanitis, A., Draganova, C., Christodoulou, C.: Comparing dif-

ferent classiﬁers for automatic age estimation. IEEE Trans. Syst.

Man Cybern. Part B: Cybern. 34(1), 621–628 (2004)

8. Geng, X., Zhou, Z.H., Zhang, Y., Li, G., Dai, H.: Learning from

facial aging patterns for automatic age estimation. In: Proceed-

ings of the 14th Annual ACM International Conference on

Multimedia, pp. 307–316. Santa Barbara, USA (2006)

9. Ailisto, H., Vildjiounaite, E., Lindholm, M., Mkel, S.M.,

Peltola, J.: Soft biometrics - combining body weight and fat

measurements with ﬁngerprint biometrics. Pattern Recognit.

Lett. 27(5), 325–334 (2006)

Soleprint Device

▶ Fingerprint, Palmprint, Handprint and Soleprint

Sensor

Soleprint Sensor

▶ Fingerprint, Palmprint, Handprint and Soleprint

Sensor

Sound

Sound is a pressure wave which is created by a vibrat-

ing object.

▶ Speech Production

Sound Generation

▶ Speech Production

▶ Voice Sample Synthesis

Sources of Evidence

▶ Sources of Information in Biometric Fusion

Sources of Information in Biometric

Fusion

ARU N ROSS

Lane Depar tment of Computer Science and Electrical

Engineering, West Virginia University, Morgantown,

WV, USA

Synonyms

Sources of evidence

Definition

Multibiometric systems rely on the evidence presented

by multiple sources of biometric information. Based on

Sources of Information in Biometric Fusion

1239