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

Подождите немного. Документ загружается.

the receiver ﬁnally reads the letter in the sender’s per-

sonal handwriting style.

Since signature is a special type of handwriting,

handwriting synthesis techniques can be adapted for

signature sample synthesis. Instead of representing

each character in handwriting synthesis, the whole

signature can be taken as a single glyph. Then the

character modeling is applied to the signature glyphs,

for example, extracting the control points or estimat-

ing the writing speed. During synthesis, the process of

character selection is applied to the learnt models of

signature glyphs. By this way, a large database of syn-

thetic signatures can be obtai ned to test the robustness

of an existing handwriting/signature veriﬁcation or

recognition method.

Summary

As an inverse biometrics problem, signatu re/handwrit-

ing sample synthesis has been studied in recent years.

Existing synthesis systems can help common users to

produce personal signature/handwriting with pleasing

visual quality. However, they do not capture all aspects

of the handwriting style. For example, the handwriting

of people may evolve gradually with their ages. It may

add more liveness to the handwriting if the effect of

time is considered. Furthermore, the research on syn-

thesis in languages other than English needs more

investigation before it becomes accessible to people in

different countries.

Related Entries

▶ Handwriting Structure

▶ Sample Synthesis from Templates

▶ Signature Features

▶ Signature Recognition

References

1. Singer, Y., Tishby, N.: Dynamic encoding of cursive handwriting.

Biometric Cybernetics 71(3), 227–237 (1994)

2. Gupta, G., McCabe, A.: A review of dynamic handwritten signa-

ture veriﬁcation. Tech. rep. (1997)

3. Oliveira, C.D., Kaestner, C., Bortolozzi, F., Sabourin, R.:

Generation of signatures by deformations In: BSDIA ’97:

Proceedings of the First Brazilian Symposium on Advances

in Document Image Analysis, pp. 283–298 Springer, London,

UK (1997)

4. Plamondon, R., Maarse, F.: An evaluation of motor models

of handwriting. IEEE Trans. Pattern Analy. Machine Intell.

19(5), 1060–1072 (1989)

5. Li, X., Parizeau, M., Plamondon, R.: Segmentation and

reconstruction of on-line handwritten scripts. Pattern Recogn.

31(6), 675–684 (1998)

6. FontGod Corperation: Handwriting fonts. http://www.fontgod.

com/ (2007)

7. Zheng, Y., Doermann, D.: Handwriting matching and its appli-

cation to handwriting synthesis. In: Proceedings of Eighth Inter-

national Conference on Document Analysis and Recognition,

vol. 2, pp. 861–865 (2005)

8. Brault, J.J., Plamondon, R.: A complexity measure of handwrit-

ten curves: modelling of dynamic signature forgery. IEEE Trans.

Syst. Man Cybernetics 23, 400–413 (1993)

9. Tappert, C.: Handwriting synthesis of a particular writer’s

style. http://csis.pace.edu/ctappert/research/researchtopics.htm

(2007)

10. Srihari, S., Cha, S., Arora, H., Lee, S.: Individuality of handwrit-

ing. J. Forensic Sci. 47(4), 856–872 (2002)

11. Lee, D.H., Cho, H.G.: A new synthesizing method for handwrit-

ing Korean scripts. Int. J. Pattern Recogn. Artif. Intelligence

12(1), 46–61 (1998)

12. Jawahar, C., Balasubramanian, A.: Synthesis of online handwrit-

ing in Indian languages. In: Proceedings of International Work-

shop on Frontiers in Handwriting Recongnition. La Baule,

France (2006)

13. Guyon, I.: Handwriting synthesis from handwritten glyphs. In:

Proceedings of Fifth International Workshop on Frontiers of

Handwriting Recognition, pp. 309–312 Colchester, England

(1996)

14. Wang, J., Wu, C., Xu, Y.Q., Shum, H.Y., Ji, L.: Learning based

cursive handwriting synthesis. In: Proceedings of Eighth Inter-

national Workshop on Frontiers of Handwriting Recognition,

pp. 157–162. Ontario, Canada (2002)

15. Wang, J., Wu, C., Xu, Y.Q., Shum, H.Y.: Combining shape and

physical models for on-line cursive handwriting synthesis. Int. J.

Doc. Analy. Recogn. 7(4), 219–227 (2005)

16. Nikitin, A.V., Popel, D.V.: Signmine algorithm for condition-

ing and analysis of human handwriting. In: Proceedings

of the International Workshop on Biometric Technologies,

pp. 179–190 Calgary, Alberta (2004)

17. Lin, Z., Wan, L.: Style-preserving English Handwriting Synthe-

sis. Pattern Recogn. 40, 2097–2109 (2007)

Signature Similarity Computation

▶ Signature Matching

1210

Signature Similarity Computation

Signature Synthesis

▶ Signature Sample Synthesis

Signature/Sign Recognition

▶ Signature Recognition

Silhouette

The set of pixels in an image corresponding to an

object of interest. In gait recognition, the object of

interest is the human subject to be recognized. These

are outlines of persons or object of interest in image s or

videos. In gait recognition, these are represented as 2D

(ﬁlled) shapes of walking persons. Most common

usage includes both the inside and the boundary of

the shape to be part of the silhouette. Silhouettes are

typically detected by subtracting the background image

from any given image, followed by some simple hole

ﬁlling and clean up operations.

▶ Gait Recognition, Silhouette-Based

Silhouette Analysis for Gait

Recognition

▶ Gait Recognition, Motion Analysis for

Similarity Metric

Any of a large number of scalars assessing the degree of

similarity, or of dissimilarity, between objects such as

biometric features or templates is similarity metric.

Examples include: correlation statistics, vector projec-

tions (cosine or inner products), and fractional Ham-

ming Distance (the fraction of bits that differ between

two bit sequences of equal length). Similarity metrics

usually obey the axioms of distances (symmetry; non-

negativity; and triangle rule inequality), but they can

incorporate nonlinear or non-Euclidean topologies

(e.g., city-block distances between points in a space

versus straight-line distances). A simil arity metric be-

tween biometric templates, after being suitably nor-

malized, is usually the input into a decision process

that renders a judgment about whether they should be

classiﬁed as ‘‘same’’ or ‘‘different.’’

▶ Palmprint Matching

▶ Score Normalization Rules in Iris Recognition

Simplifying Passenger Travel

Program

Synonyms and Acronyms

ABG; CANPASS; IRIS; NEXUS; Privium; SPT

Definition

Organized by IATA (the international air transport

association representing 230 airlines), the simplify ing

passenger travel initiative refers to a combination of

programs agreed by airlines, airports, and governmen-

tal regulatory bodies for the purpose of simplifying the

experience of air travel and making it more efﬁcient.

Not focused solely on expediting secu rity procedures

for departing passengers, SPT programs also incl ude

the use of biometrics for automated immigration

clearance for arriving passengers, streamlined border-

crossing, automated check-in, and enhancing other

airport processes. National programs that use iris rec-

ognition for these purposes include IRIS (UK): Iris

recognition immigration system; Privium (NL); ABG

(Germany); and the Canadian and US border-crossing

programs, CANPASS and NEXUS.

▶ Score Normalization Rules in Iris Recognition

Simplifying Passenger Travel Program

1211

Simultaneous Capture of Iris and

Retina for Recognition

DAV I D USHER,YASUNARI TOSA,MARC FRIEDM AN

Retica Systems, Inc., Waltham, MA, USA

Synonyms

Iris retina biometric fusion

Definition

The simultaneous imaging of the iris and retina for the

purposes of biometric recognition. A specialized optical

engine is used to illuminate and image both biological

features. Acquisition software extracts optimal images

from acquired video sequences. The resultant digital

images are analyzed and biometric information is

extracted and stored. Information from both the iris

and retina is combined or fused for the purposes of

verifying or determining the identity of the individual.

The ﬁxed anatomical proximity of the iris and retina

facilitate their simultaneous capture by a single device.

Combining the iris and retina traits aims to offer an

advantage over unimodal iris and retina biometric

systems.

Introduction

Within the ﬁeld of biometrics, the iris has received

considerable attention. Iris biometric system s are

commercially available and research is expandi ng.

Only recently has the technology begun to meet its

commercial potential. The reasons for this involve

meeting the many challenges necessary to create sys-

tems that combine high quality imag ing, good human

factors engineering, and high quality software algo-

rithms for image capture, segmentation, encoding,

and matching. Retina-based identiﬁcation has long

been perceived as a robust biometric solution, but

very few practical applications or commercially viable

products have been demonstrated. EyeDentify Inc. de-

veloped a retinal biometric product [1, 2] that demon-

strated reasonable performance [3]. However, it

suffered from a perception that its human interface

was intrusive. Optibrand Ltd. developed a retinal

biometric device for the livestock market [4]. More

recently, Retica Systems Inc. has developed a biometric

acquisition system that combines the retina with iris

biometrics [5]. As biometric systems attempt to meet

the demands of real world applications, multibio-

metric systems are receiving considerable attention. It

is believed that some of the limitations imposed by

unimodal systems can be overcome by using multiple

biometric modalities [6]. Both iris and retina systems

can suffer from problems associated with unimodal

systems. These include noisy data, intraclass variation,

nonuniversality, and susceptibility to spoof attacks. Al-

though iris technologies have demonstrated high levels

of performance, research with the aim of mitigating

poor iris image quality is ongoing, and attempts have

been made to improve performance by combining the

iris with other biometric traits including ﬁngerprint [7]

and face [8 ]. The ﬁxed anatomical proximity of the iris

and retina facilitates their simultaneous capture using

a single system. Biometric traits are best combined

when their discriminating power is evenly balanced

and their content is independent. The topology of the

retinal blood vessels is indepe ndent of the texture of

the iris. It therefore may be possible to improve bio-

metric performance by combining balanced iris and

retina recognition technologies into a single device.

Anatomical Background

The eye can be divided into the anterior and posterior

segments, Fig. 1. The iris is found in the anterior

segment that also includes the cornea and lens.

The iris is constructed of pigmented ﬁbrovascular tis-

sue layered onto a back surface of pigmented epitheli-

um cells. Crypts and freckles add to the observed

pattern. The texture variation across the iris is distinc-

tive and it is this information that is encoded forming

the iris biometric signal. The retina is found in the

posterior segment that comprises the back two-thirds

of the eye. Light is refracted by the cornea and lens

through the pupil onto the retina, a thin layer of neural

cells that lines the interior surface of the eye (the

fundus). The fundus, as seen using a digital fundus

camera, is shown in Fig. 2 where the images of the left

eyes of two identical twins are shown. The retinal

blood vessel pattern is the subject of biometric encod-

ing methods. Two major blood vessel systems supply

the retina. The outer retinal layers are supplied by a

1212

Simultaneous Capture of Iris and Retina for Recognition

choroidal blood vessel network. The choroidal vessels

form a grid-like pattern and are not generally visible

using standard digital fundus cameras and refractive

ophthalmoscopes. The inner layers of the retina are

supplied by the

▶ central retinal artery. There is also

one main collecting trunk, the

▶ central retinal vein.

These two blood vessels form

▶ bifurcations as they

emerge from the optic disc and branch out through the

nerve ﬁber and ganglion cell layers forming an extend-

ed network throughout the retina. The optic disc is the

point where the optic nerve breaks out into the retina

and can be seen as the bright spots in Fig. 2. Several

studies have concluded that the bran ching patterns of

the retinal arterial and venous systems have character-

istics of a fractal [9, 10]. It has been suggested that a

nonequilibrium Laplacian process could be involved in

▶ retinal angiogenesis [10] and that ﬂuctuations in the

distribution of embryonic cell-free spaces provide the

randomness needed for fractal behavior and for

the uniqueness of each individual’s retinal vascular

pattern [9]. This fractal-like growth occurs in the em-

br yonic stages of humans and provides for uniqueness

even in the case of identical twins, Fig. 2. The anato-

mical stability of the iris and retina biometric traits

must be considered. While it is expected that both

suffer from minimal normal age-related changes,

both are effected by various disease states. Large-scale

studies speciﬁcally addressing the stability of the iris,

especially for biometrics, have yet to be performed

[11]. This is also the case for retinal biometrics.

Challenges

The principal challenges for an ocular imaging biomet-

ric device broadly fall into three categories encompass-

ing the imaging system, its human interface, and

software analysis algorithms:

1. Imaging system. The challenge for the imaging sys-

tem is to record stable images that best illuminate

biometric features. In the case of the iris, standards

Simultaneous Capture of Iris and Retina for

Recognition. Figure 1 Schematic diagram of the human

eye (right). (Courtesy of National Eye Institute, National

Institutes of Health).

Simultaneous Capture of Iris and Retina for Recognition. Figure 2 Fundus camera images corresponding to the left

eyes of two identical twins. Retinal blood vessels can be seen to form a branching network centered on the optic disc.

The optic disc can be seen as a bright spot near the center of each image. Comparing the two images, the topology of the

blood vessels is notably different. (Courtesy of Prof. Michael Larsen, Glostrup Hospital, University of Copenhagen).

Simultaneous Capture of Iris and Retina for Recognition

1213

[12] dictate two basic requirements: (1) near-infra-

red illumination must be used, (2) pixel and spatial

resolution limits must be met. Illumination across

the eye must produce an image with even levels of

contrast throughout and with clearly deﬁned iris

boundaries. Iris texture should be emphasized

and the optical system must minimize obscur-

ing reﬂections incl uding

▶ Purkinje images and

those from ambient light sources. In the case of the

retina, the interior surface of the eye must be imaged

through the refracting surfaces of the cornea and lens

and through its natural aperture, the pupil. A suitable

ﬁeld-of-view containing a high level of blood vessel

detail is required. The blood vessels must show a

suitable level of contrast. Optical appliances such as

glasses, contact lens and other types of face guards,

and masks add additional imaging challenges for

both iris and retina systems, as they can contain

scratches and generally exhibit poor transmission

properties. In addition, in the case of a dual iris-

retina system, iris and retina imaging systems should

complement each other and steps must be taken such

that there is no interference between the two.

2. Human interface. Human interface challenges en-

compass ﬁxation and targeting, illumination con-

siderations, and distance requirements, and are

highly dependent on the application of the tech-

nology. In general, a biometric device should in-

convenience the user as little as possible while

facilitating repeatable and stable imaging of the

biometric traits. The challenges associated w ith

imaging the retina dictate that, at least in the short

term, a passive imaging system with a level of

active participation by the user is required. If a

passive system is employed, then a suitable align-

ment tool is required. This ﬁxture must be straight-

forward and intuitive to use and be capable of

aligning the user to a deﬁned degree of accuracy.

3. Software analysis. Software analysis challenges

involve live acquisition, feature extraction, encoding,

and matching. The task of the image acquisition step

is to identify from video sequences if any acceptable

views of the biometrics have been presented and, if

so, to extract and record the best examples. In this

context the term best means an image from which

the biometr ic signal can be encoded with the high-

est degree of accuracy. Image acquisition methods

must process continuous video sources. Processing

constraints are therefore high. A deﬁnition of

image quality is needed and thresholds must be

applied to exclude unacceptable images. There is

an inherent trade-off between the image quality

thresholds applied during image acquisition, the

human interface, and the subsequent efﬁcacy of

the encoding. Image quality constraints set too

high may result in a more prolonged and difﬁcult

user experience or ubiquitous failures-to-acquire.

Image quality constraints set too low may compro-

mise encoding and therefore potentially degrade

matching performance. Feature extraction techni-

ques are used to reduce the acquired images into

biometric signals. Methods must accurately extract

the unique features present in the image and efﬁ-

ciently encode them to facilitate matching. For iris

analysis systems, the iris must ﬁrst be located and

separated from the rest of the image. This necessar-

ily includes identify ing areas of iris occlusion. In

the case of the retina, the blood vessel network

must be separated from other features within the

retinal images. Encoding methods must provide an

efﬁcient characterization of the biometric features

that facilitates accurate and rapid matching meth-

ods. Finally, matching algorithms must deﬁne a

similarity score such that scores from pairs of sig-

nals from the same indiv idual show a high separa-

tion from scores generated from signals of different

individuals. Operational constraints often dictate

that matching must be rapid.

Image Quality Considerations

Systems that combine a good image capture hardware

design with a straight forward human interface can still

suffer from large variations in image quality. Inherent

anatomical, behavioral, and environmental variations

introduce confounding factors. A large range in iris

reﬂectance can result in low contrast for one of its

boundaries (e.g., dark irises can exhibit poor contrast

with the pupil. Highly reﬂective irises can result in a

low level of contrast for the iris–sclera boundary).

Variations in the reﬂectivity of the retina can result

in a range of retinal image brightness and contrast.

Variations in the optical efﬁciency of the eye and the

variable use of glasses or contact lenses can result in a

range of achievable focus for a ﬁxed focused optical

system imaging the retina, and add unwanted reﬂec-

tions to iris images. Ambient light sources also affect

1214

Simultaneous Capture of Iris and Retina for Recognition

the size of a user’s pupil. In addition to altering the

appearance of the iris, pupil scale changes can dyna-

mically affect the illumination on the retina. Variations

in user alignment can introduce eye-gaze to iris images

that can adversely affect segmentation and can intro-

duce nonline ar transformations to the blood vessel

pattern as projected onto a two dimensional retinal

image. Standards for iris image quality have been de-

ﬁned [12]. No such standard exists for retinal images.

Simultaneous Capture of the Iris and

Retina

This section discusses Retica’s fusion of its iris

and retina technologies for simultaneous capture by a

single-presentation biometric device. Solutions to the

challenges discussed earlier combine to form a novel

iris–retina imaging system, its associated human inter-

face, and proprietary software analysis algorithms for

automated acquisition, encoding, and matching.

A more detailed description of Retica’s dual iris–retina

technolog y can be found in [5].

1. Imaging system. Proprietary iris and retina optical

systems are arranged in a single housing. Both

systems employ standard Video Graphics Array

(VGA) resolution cameras. In the iris optical sys-

tem, illumination is provided by a set of narrow-

band near-infrared (NIR) LED’s. A ﬁxed focus

system with a focal length of 27 cm is used. Care

was taken to meet the iris image quality standards

[12] for resolution, contrast, and noise levels. The

resolution of the iris imaging system is 25 pixels per

mm. An example of an iris image recorded using

Retica’s imaging system is shown in Fig. 3(a).

The retinal optical system was simpliﬁed by using

a10

∘

ﬁeld-of-view. This relatively narrow ﬁeld

allowed image capture from a greater distance.

The retina camera was set at a horizontal angle of

15.5

∘

and 1.5

∘

below line of sight, (see Fig. 4). At

this angle its ﬁeld-of-view was centered on the

optic disc region. This region was chosen, as it has

a high concentration of blood vessels. In addition,

the optic disc is close to the pivot point for eye

rotation. It is therefore the most stable part of the

retina in terms of transformations in the recorded

images as a result of eye movements. Illumination

is provided by a narrowband NIR LED reﬂected off

a beamsplitter. An aper ture is imaged by a large lens

to a 2 mm spot just before the cornea uniformly

illuminating the optic disc region. The resolution

of the retina imaging system is 90 pixels per mm.

Optimal positioning is approximately 27 cm from

the front panel. Fig ure 3(b) shows an example of a

captured image of the retina. The optic disc can be

seen near the center of the image along with radiat-

ing blood vessels.

2. Human interface. A beam splitter placed in front of

the iris camera (Fig . 4) is used to create a targeting

system for the user consisting of a colored illumi-

nating ring surrounding a colored disk containing

Simultaneous Capture of Iris and Retina for Recognition. Figure 3 (a) An example of an iris image captured using

Retica’s iris optical system and associated software analysis tools. The field-of-view comprises 26 by 20 mm encompassing

the whole visible section of the eye. (b) An example of a retinal image captured using Retica’s retinal optical system and

associated software analysis tools. A 10

∘

ﬁeld-of-view was centered approximately on the optic disc).

Simultaneous Capture of Iris and Retina for Recognition

1215

crosshairs. When the subject is able to see the

disk centered within the ring, they are in correct

lateral alignment. The user is then asked to move

forward maintaining lateral alignment. In doing so

they move through the ﬁxed focus of the iris system

and across the optical axis of the retina system.

Empirical evidence showed that lateral alignment

could be achieved easily even for inexperienced

users. However, there was a large range in the

expectation of the distance positioning t hat was

required. Visual and auditory cues are used to

guide the user as they move towards the device.

An optical rangeﬁnder is used to set the color of

the targeting system’s colored disk to green when

the user is at approximately the correct distance.

In addition to this, images meeting predeﬁned

quality thresholds (see the following section) trig-

ger an audible tone. The user is instructed to con-

tinue mov ing towards the device until the tone

stops, then to move slowly backwards again moving

through the region coinciding with the tone once

more. These steps are repeated until the acquisition

process is interrupted when recorded data have met

predeﬁned thresholds for quality and quantity. The

user must move through the optimal alignment

position; they are not required to hol d a ﬁxed

position. Moderately experienced users could be

acquired in less than 1 second. Most inexperienced

users were acquired in less than 15 seconds. The

active cooperation of the user combined with an

intuitive alignment tool reduce the effect of eye

gaze in the acquired iris images.

3. Software analysis. Iris and retina acquisition

algorithms automate the acquisition process.

Image quality thresholds for iris and retina image

focus are deﬁned along with measures of iris occlu-

sion and retinal blood vessel content. A video

rate assessment of image quality is then used to

select best iris and retina frames as the user passes

though optimal alig nment. Acquired images are

passed on to proprietary iris and retina encoding

algorithms. Encoded templates are then matched

against existing databases.

Conclusion

The section introduced the concept of simultaneous

capture of iris and retina for biometric recognition.

Some of the challenges associated with iris and retina

biometric systems were discussed. A bimodal system

that demonstrated that the iris and retina can be ac-

quired simultaneously was outlined and represents a

unique contribution to the ﬁeld of biometrics. Although

unimodal results are good, intraclass variations can

present problems for both biometrics. As discussed

earlier, anatomical, behavioral, and environmental fac-

tors can result in a range in image quality for both the

iris and retina. The level of success of a recognition

system can be largely deﬁned by how well its encoding

and matching methods are able to manage these varia-

tions. For example, inexact iris localization because of

poor iris image quality or failures in blood vessel seg-

mentation because of poor retinal image quality ulti-

mately result in poor genuine match scores.

Commercially available iris recognition systems exhibit

nonzero error rates. Problems such as occlusion, motion

blur, specular reﬂections, and pose contribute to

Simultaneous Capture of Iris and Retina for Recognition. Figure 4 Outline of Retica’s dual retina-iris optical system.

1216

Simultaneous Capture of Iris and Retina for Recognition

intraclass variations. It may be possible to enhance

performance by using a dual iris–retina system. The

topology of the retinal blood vessel pattern is completely

uncorrelated to the texture patterns on the iris. They

therefore represent complementary sources of informa-

tion. While it is true that they share some ubiquitous

failures, (e.g., someone with no eye has neither an iris

nor a retina) various obfuscatory factors affecting the iris

or the retina are either uncorrelated or anticorrelated.

For example, eyelid and eyelash occlusion has no rela-

tion to retinal blood vessel detail. Highly dilated pupils

that can cause problems for iris systems aid imaging of

the retina. Both traits represent strong biometrics po-

tentially facilitating a more balanced fusion than the

combination of a strong biometric with a weaker one.

Both biometric features are enclosed organs and cannot

be altered without endangering vision. However, there

is a risk of spoof iris attacks, and antispooﬁng measures

are being actively investigated. Adding a requirement

for retinal identiﬁcation signiﬁcantly increases the chal-

lenge of hoax enrollment. Next generation iris

acquisition systems aim to relax constraints imposed

on users in terms of capture volume, standoff distance,

and motion. While the system proposed by Retica

demonstrated straightforward retina acquisition at

27 cm, less-constrained retina acquisition presents a

signiﬁcant challenge. Dual iris–retina capture systems

are therefore likely to be restricted to applications that

require the highest accuracy, for cooperative users,

with relatively constraining human interfaces.

Related Entries

▶ Eye Features and Anatomy

▶ Iris Acquisition Device

▶ Iris Image Data Interchange Formats, Standardi-

zation

▶ Retina Recognition

▶ Vein and Vascular Recognition, Overview

References

1. Johnson, J.C., Hill, R.B.: Eye fundus optical scanner system and

method. US Patent No. 5668842 (1990)

2. Hill, R.B.: Retina identiﬁcation. In: A.K. Jain, R. Bolle,

S. Pankanti (eds.) Biometrics Personal Identiﬁcation in Net-

worked Society., chap. 6, pp. 123–142. Springer, London (1999)

3. Holmes, J.P., Wright, L.J., Maxwell, R.L.: A performance evalua-

tion of biometric identiﬁcation devices. Tech. Rep. SANDIA91–

0276, Sandia National Laboratory (1991)

4. Golden, B.L., Rollin, B.E., Switzer, R.V., Comstock, C.R.: Retinal

vasculature image acquisition apparatus and method. US Patent

No. 6766041 B2 (2004)

5. Usher, D., Tosa, Y., Friedman, M.: Ocular biometrics: Simulta-

neous capture and analysis of the retina and iris. In: N. Ratha,

V. Govindaraju (eds.) Advances in Biometrics: Sensors, Algo-

rithms and Systems, chap. 8, pp. 133–155. Springer, London

(2008)

6. Ross, A., Jain, A.J.: Multimodal biometrics: An overview. In:

Proceedings of 12th European Signal Processing Conference

(EUSIPCO), pp. 1221–1224. Vienna, Austria (2004)

7. Nandakumar, K., Chen, Y., Dass, S.C., Jain, A.K.: Quality-based

score level fusion in multibiometric systems. In: Proceedings

of International Conference on Pattern Recognition (ICPR),

pp. 20–24. Hong Kong (2006)

8. Wang, Y., Tan, T., Jain, A.K.: Combining face and iris biome-

trics for identity veriﬁcation. In: Proceedings of four th

International Conference on Audio- and Video-Based Biometric

Person Authentication (AVBPA), pp. 9–11. Guildford, UK (2003)

9. Mainster, M.A.: The fractal properties of retinal vessels: Embry-

ological and clinical implications. Eye 4(1), 235–241 (1990)

10. Masters, B.R.: Fractal analysis of the vascular tree in the human

retina. Annu. Rev. Biomed. Eng. 6, 427–452 (2004)

11. Wildes, R.P.: Iris recognition. In: A.K. Jain, D. Maltoni, D. Maio

(eds.) Biometric systems. Technology, design and performance

evaluation, pp. 63–95. Springer, London (2004)

12. ANSI/INCITS: American National Standard For Informa-

tion Technology. Iris image interchange format. ANSI INCITS

379–2004 (2004)

Skilled Forgery

In signature veriﬁcation, skilled forgeries represent the

imitations that are intentionally performed, that is,

where the forger actively tries to imitate the signature

from another user.

▶ Signature Databases and Evaluation

Skin Classification

▶ Skin Detection

Skin Classification

1217

Skin Color Detection

▶ Skin Detection

Skin Detection

AHMED ELGAMMAL,CRY S TAL MUANG,DUNXU HU

Department of Computer Science, Rutgers University,

Piscataway, NJ, USA

Synonyms

Skin classiﬁcation; Skin color detection

Definition

Skin detection is the process of ﬁnding skin-colored

pixels and regions in an image or a v ideo. This process

is typically used as a preprocessing step to ﬁnd regi-

ons that potentially have human faces and limbs in

images. Several computer vision approaches have been

developed for skin detection. A skin detector typically

transforms a given pixel into an appropriate color

space and then use a skin classiﬁer to label the pixel

whether it is a skin or a nonskin pixel. A skin classiﬁer

deﬁnes a decision boundary of the skin color class

in the color space based on a training database of

skin-colored pixels.

Introduction

Skin color and textures are important cues that people

use consciously or unconsciously to infer variety of

culture-related aspects about each other. Skin color

and texture can be an indication of race, health, age,

wealth, beauty, etc. [3]. However, such interpretations

vary across cultures and across the history. In images

and videos, skin color is an indication of the exis-

tence of humans in such media. Therefore, in the last

two decades extensive research has focused on skin

detection in images. Skin detection means detecting

image pixels and regions that contain skin-tone color.

Most of the research in this area has focused on detect-

ing skin pixels and regions based on their color. Very

few approaches attempt to also use texture informa-

tion to classify skin pixels.

As will be described shortly, detecting skin pixels is

rather a computationally easy task and can be done

very efﬁciently, a feature that encourages the use of

skin detection in many video analys is applications. For

example, in one of the early applications, detecting

skin-colored regions was used to identify nud e pictures

on the internet for the sake of content ﬁltering [6]. In

another early application, skin detection was used to

detect anchors in TV news videos for the sake of video

automatic annotation, archival, and retrieval [1]. In

such an application, it is typical that the face and the

hands of the anchor person are the largest skin-tone

colored region in a given frame since, typically, news

programs are shot in indoor controlled environments

with man-made background materials that hardly con-

tain skin-colored objects. In many similar applications,

where the background is controlled or unlikely to

contain skin-colored regions, detecting skin-colored

pixels can be a very efﬁcient cue to ﬁnd human faces

and hands in images. An example in the context of

biometrics is detecting faces for face recognition in an

controlled environment.

Detecting skin-colored pixels, although seems a

straightforward easy task, has proved quite challenging

for many reasons. The appearance of skin in an image

depends on the illumination conditions (illumination

geometry and color) where the image was captured.

Humans are very good at identifying object colors in a

wide range of illuminations, this is called

▶ color con-

stancy. Color constancy is a mystery of perception.

Therefore, an important challenge in skin detection is

to represent the color in a way that is invariant or at

least insensitive to changes in illumination. As will be

discussed shortly, the choice of the color space affects

greatly the performance of any skin detector and its

sensitivity to change in illumination conditions. An-

other challenge comes from the fact that many objects

in the real world might have skin-tone colors. For

example, wood, leather, skin-colored clothing, hair,

sand, etc. This causes any skin detector to have many

false detections in the background if the environment

is not controlled.

1218

Skin Color Detection

A Framework for Skin Detection

Skin detection process has two phases: a training phase

and a detection phase. Training a skin detector involves

three basic steps:

1. Collecting a database of skin patches from different

images. Such a database typically contains skin-

colored patches from a variety of people under

different illumination conditions.

2. Choosing a suitable color space.

3. Learning the parameters of a skin classiﬁer.

Given a trained skin detector, identi fying skin

pixels in a given image or video frame involves:

1. Converting the image into the same color space

that was used in the training phase.

2. Classifying each pixel using the skin classiﬁer as

either a skin or a nonskin.

3. Typically post processing is needed using morphol-

ogy to impose spatial homogeneity on the detected

regions.

In any given color space, skin color occupies a part

of such a space, which might be a compact or large

region in the space. Such region is usually called the

skin color cluster. A skin classiﬁe r is a one-class or two-

class classiﬁcation problem. A given pixel is classiﬁed

and labeled whether it is skin or nonskin given a model

of the skin color cluster in a given color space. In the

context of skin classiﬁcation, true positives are skin

pixels that the classiﬁer correctly labels as skin. True

negatives are nonskin pixels that the classiﬁer correctly

labels as nonskin. Any classiﬁer makes errors: it can

wrongly label a nonskin pixel as skin or a skin pixel as a

nonskin. The former type of errors is referred to as

false positives (false detections) while the later is false

negatives. A good classiﬁer should have low false posi-

tive and false negative rates. As in any classiﬁcation

problem, there is a tradeoff be tween false positives and

false negatives. The more loose the class boundary,

the less the false negatives and the more the false

positives. The tighter the class boundary, the more

the false negatives and the less the false positives. The

same applies to skin detection. This makes the choice

of the color space extremely important in skin detec-

tion. The color needs to be represented in a color space

where the skin class is most compact in order to be able

to tightly model the skin class. The choice of the color

space directly affects the kind of classiﬁer that should

be used.

Skin Detection and Color Spaces

As was highlighted by Forsyth and Fleck [6] the human

skin color has a restricted range of hues and is not

deeply saturated, since the appearance of skin is

formed by a combination of blood (red) and melanin

(brown, yellow). Therefore, the human skin color does

not fall randomly in a given color space, but clustered

at a small area in the color space. But it is not the same

for all the color spaces. A variety of color spaces has

been used in skin detection literature with the aim of

ﬁnding a color space where the skin color is invariant

to illumi nation conditions. The choice of the color

spaces affects the shape of the skin class, which affects

the detection process. Here, some color spaces, which

are typically used in skin detection, are brieﬂy de-

scribed, and the way they affect the skin detection is

discussed. The goal of the discussion is to hig hlight

answers to the followi ng questions: Given a skin patch,

where will it be located in a given color space? Given a

skin patch, what effect will chan ging the illumination

intensity have in its location in a given color space?

Given skin patches from different people from the

same race, how are all these patches related in a given

color space? Given skin patches from different people

races, how are all these patches related in a given color

space? Figures 1 and 2 help illustrate the answers for

these questions. Figure 1 shows density plots for skin-

colored pixels obtained from images of different Asian

people plotted in different color spaces. Figure 2 shows

density plots for skin-colored pixels from different

people from different races: Asian, African, and Cau-

casian plotted in different color spaces.

RGB Color Spac e and Skin Detection: RGB color

space is the most commonly used color space in digital

images. It encodes colors as an additive combination of

three primary colors: red (R), green (G), and blue (B).

RGB Color space is often visualized as a 3D cube where

R, G, and B are the three perpendicular axes. One main

advantage of the RGB space is its simplicity. However,

it is not perceptually uniform, which means distances

in the RGB space do not linearly correspond to human

perception. In addition, RGB color space does not

separate

▶ luminance and ▶ chrominance, and the R,

Skin Detection

1219