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

Подождите немного. Документ загружается.

Human detection algorithms can also be differen-

tiated based on whether or not they utilize motion

cues. Humans have movement patterns that are dis-

tinct from other object classes, and some algorithms

exploit motion features such as periodicity and motion

symmetry to detect humans in video [7]. While

motion provides a very powerful cue to detect humans,

it is important for human detection algorithms to

utilize both appearance and motion features in a bal-

anced manner, so as to detect people irrespective of

whether they are moving or stationary [2, 4, 8].

Typically human detection algorithms aim at

providing only the location and scale of the people in

the scene. More recently there has been a growing

emphasis on additionally recovering the silhouette

shape of each detected person. Such algorithms typi-

cally require training images in which the silhouettes

are manually annotated in each image [9], though

there are other algorithms that require either only

very limited or no manual annotation of the training

dataset [8]. Since humans have articulated limbs, can

occur in different poses, and may carry objects (e.g.,

briefcase, backpack, etc.), acquiring the actual silhou-

ette shape can provide valuable cues for gait and activ-

ity recognition systems.

Human Tracking

Tracking is the process of temporally associating the

location of the target object (e.g., human) from one

frame to the next in a sequence of images (v ideo).

A detailed survey on object tracking can be found in

[10]. Human tr acking is of considerable interest to

security applications, and the tracking information is

useful for determining the orig in, pathway, and desti-

nation of each person (Fig. 2). As most human track-

ing approaches employ standard (generic) object

tracking algorithms, an overview of these techniques

is presented here.

Most tracking algorithms can be broadly classiﬁed

into two categories: (1) data association and ﬁltering

techniques or (2) target representation and localization

methods. The ﬁrst category is a class of solutions

characterized by associating a set of detection locations

from the previous image to a set of detections in the

current image. The second category is distinguis hed

by building a target model of the appearance of

the object in the initial frame and searching for the

object location in successive frames by generating can-

didate models and ﬁnding the best possible target-

candidate match. The tracking task is complicated by

Human Detection and Tracking. Figure 2 Trajectories of tracked people.

710

Human Detection and Tracking

changes in illumination, image noise, nonrigid target

motion, and occlusion. For practical applications,

algorithms from both categories may be combined to

obtain robust and efﬁcient trackers.

Data Association and Filtering

Tracking techniques based on association assume that

there is a set of detection locations in an image at time

t and a set at time tþ1. The goal of these trackers is to

ﬁnd a mapping from the set at time t to the subsequent

set at time tþ1, establishing detection correspondence

between the frames. The simplest solution is to map

the set at time t to its successor according to Euclidean

distance. A more complex approach may incorporate

the velocity of each detection into the computation,

assuming both positional proximity and smooth con-

tinuity in the movement [11]. These data association

techniques use simple assumptions and minimize

some criteri a according to kinematic features extracted

from the video sequence. Naturally occlusion is a sig-

niﬁcant challenge with these techniques, as the map-

ping is more difﬁcult to recover when the cardinalities

of the two sets are not equal (i.e. , missing data).

Filtering strategies improve upon the basic data

association approach by supplying a speciﬁc model of

object motion and algorithms by which one can re-

cover the optimal sequence of locations of the object

throughout the video. These strategies generally follow

a predict and update framework, where the new posi-

tion and velocity of the object at time tþ1 is hypothe-

sized/predicted by the model from time t, and then the

model is updated after the frame at time tþ1 is pro-

cessed. These me thods tend to smooth the trajectories

(removing noise) and improve the robustness to oc-

clusion (as new object locations are hypothesized even

when data are not present).

One of the simplest and most common of these

ﬁltering approaches is the Kalman ﬁlter [12]. The

Kalman ﬁlter is a state-space model that makes

assumptions that the position and velocity at time

tþ1 is a linear function of the position and velocity

of the object at time t plus some additive Gaussian

noise and that the posterior distribution of the hidden

state variable is Gaussian. These assumptions allow the

model to recover the optimal positions and velocities

from the data locations (assuming the object motion

obeys the model assumptions). Unfortunately

trajectories of human motion can at times be nonline-

ar, and in such cases the Kalman ﬁlter framework is

likely to fail. Extended Kalman ﬁlters have been pro-

posed to relieve the linear model assumption by taking

the ﬁrst-order approximation of a non-linear process

and casting the problem in a similar predict-update

framework [13].

Particle ﬁltering is a general ﬁltering strategy

that relieves the previous Gaussian assumption, allow-

ing the posteri or distribution to be generic. This is

achieved by representing the posterior distribution

as a set of weighted points (particles) in the state

space. This representation is predicted at time tþ1

by sampling an importance density function (which

approximates the posterior) at time t. It is then updated

at each frame by changing the weight of each sample,

thus allowing the modes of the density to propagate

by assimilating new observations. This type of ﬁlter-

ing was populari zed by [14], where par ticle ﬁlters

were employed to track objects through cluttered

scenes.

The drawback to most of these algorithms is that in

most cases the models do not incorporate appearance

information into the system, relying strictly on kinematic

data (position and velocity of the detections). How-

ever, there are many examples where a tracker that

could take advantage of appearance information

would signiﬁcantly increase the tracking performance,

for instance when tracking a person wearing a red shirt

walking in front of a white wall. The next category of

trackers addresses this type of tracking.

Target Representation and Localization

Target representation and localization tracking frame-

works initially extract an appearance model of the

target object from the ﬁrst image, and then search a

series of candidate locations in the next image to ﬁnd

the best matching candidate. Unlike the data associa-

tion and ﬁltering methods, this category of tracking

algorithms is mostly a bottom-up process, and essen-

tially performs detection in each frame. The algorithms

must be able to handle appearance changes of the

target over time as it is being tracked.

Popular features used for target representation incl-

ude color histograms, image gradients, and covariance

matrices (of features) within an image patch contain-

ing the target object. Localization techniques varying

Human Detection and Tracking

711

from global exhaustive search to local heuristic opti-

mizations have been employed to search for the target

locations in subsequent image frames. Many localiza-

tion algorithms beneﬁt from exploiting a spatio-

temporal locality constraint, which assumes that the

location of the object from one frame to the next

changes gradually. This constraint helps reduce the

target search space and can thus result in faster, real-

time tracking algorithms.

Mean shift, a nonparametric density gradient esti-

mator [15], is a common method used to track objects

by ﬁnding the mode (peak) of the similarity surface

generated by comparing the object appearance model

(e.g., color histogram) w ith the target candidates. Sim-

ilarity is evaluated as the Bhattacharyya coefﬁcient

between the model and candidate distributions. This

algorithm performs a local optimization on the search

surface starting from the previously known object lo-

cation and is well known for its computational efﬁ-

ciency (real-time frame rates can be obtained). Robust

tracking results are obtained under variable environ-

mental (illu mination, occlusions), object (articulate,

nonlinear motion), and camera (static, moving, jitter)

conﬁgurations.

The use of covariance features for the target repre-

sentation was proposed by [3]. The covariance matrix

of features (e.g., position, intensity, color, gradients)

extracted from an image patch enables a compact

representation of both the spatial and the statistical

properties of the object. The tracker performs a search

in the image by comparing the given covariance model

with the covariance matrix at each possible location

using an appropriately deﬁned distance metric. The

location which is most similar to the target model is

assigned to be the new target position in the image.

Summary

Human detection algorithms detect the presence of

people in imagery and must accommodate all of the

appearance variations while not selecting non‐human

entities. Human tracking temporally associates the

human detections within video sequences to generate

trajectories, and is complicated by short- and long-

term occlusions. By providing information regarding

the location and movement of humans in the scene,

human detection and tracking algorithms enable

applications such as remote biometrics and video

surveillance.

Related Entries

▶ Image Pattern Recognition

References

1. Oren, M., Papageorgiou, C., Sinha, P., Osuma, E., Poggio, T.:

Pedestrian detection using wavelet templates. In: Proceedings of

Computer Vision and Pattern Recognition (1997)

2. Dalal, N., Triggs, B., Schmid, C.: Human detection using orien-

ted histograms of ﬂow and appearance. In: Proceedings of

European Conference on Computer Vision (2006)

3. Tuzel, O., Porikli, F., Meer, P.: Human detection via classiﬁcation

on riemannian manifolds. In: Proceedings of Computer Vision

and Pattern Recognition (2007)

4. Viola, P., Jones, M., Snow, D.: Detecting pedestrians using

patterns of motion and appearance. In: Proceedings of

International Conference Computer Vision (2003)

5. Wu, Y., Yu, T.: A ﬁeld model for human detection and

tracking. IEEE Trans. Patt. Analy. and Mach. Intell. 28(5),

753–765 (2006)

6. Ramanan, D., Forsyth, D., Zisserman, A.: Strike a pose: Tracking

people by ﬁnding stylized poses. In: Proceedings of Computer

Vision and Pattern Recognition (2005)

7. Lee, S., Liu, Y., Collins, R.: Shape variation-based frieze pattern

for robust gait recognition. In: Proceedings of Computer Vision

and Pattern Recognition (2007)

8. Sharma, V., Davis, J.: Integrating appearance and motion cues

for simultaneous detection and segmentation of pedestrians. In:

Proceedings of International Conference Computer Vision (2007)

9. Leibe, B., Seemann, E., Schiele, B.: Pedestrian detection in

crowded scenes. In: Proceedings of Computer Vision and Pat-

tern Recognition (2005)

10. Yilmaz, A., Javed, O., Shah, M.: Object tracking: A survey. ACM

Comput. Surv. 38(4) (2006)

11. Rangarajan, K., Shah, M.: Establishing motion correspondence.

Comp. Vis. Graph. Img. Proc. 54(1), 56–73 (1991)

12. Kalman, R.: A new approach to linear ﬁltering and prediction

problems. Trans. ASME-J. Basic Eng. 82, 35–45 (1960)

13. Julier, S., Uhlmann, J.: A new extension to the kalman ﬁlter

to nonlinear systems. In: SPIE AeroSense Symposium (1997)

14. Isard, M., Blake, A.: Condensation – conditional density

propagation for visual tracking. Int. J. Comp. Vis. 29(1),

5–28 (1998)

15. Comaniciu, D., Ramesh, V., Meer, P.: Kernel-based object

tracking. IEEE Trans. Patt. Analy. and Mach. Intell. 25(5),

564–577 (2003)

Human Factors

▶ Ergonomic Design for Biometric Systems

712

Human Factors

Human Movement, Psychology

Synonyms

Action Categorization; Action Understanding

Definition

The psychology of human movement is a broad rang-

ing ﬁeld that includes both how the motor control

system produces movements, and how the sensory

system perceives these movements itself and from

others. Since both the structure of the body and the

strategy for producing movements are unique they

provide constraints that are potentially important for

the sensory interpretation of movement. Applied areas

of study in the psychology of human movement in-

clude sports psychology and social psychology, partic-

ularly when it applies to nonverbal communication

interpretation of visual information from movements

such as gait is of particular interest for biometrics.

In the do main of visual perception, the psychology

of human movement perception is becoming an

increasingly important example of how the visual sys-

tem processes a complex signal changing over time and

attaches meaning and social signiﬁcance to this signal.

▶ Psychology of Gait and Action Recognition

Human-Biometric Sensor Interaction

(HBSI)

▶ Ergonomic Design for Biometric Systems

Human–Computer Interaction (HCI)

and User Interfaces

Human–Computer interaction is the command and

information ﬂow that streams between the user and

the computer. It is usually characterized in terms of

speed, reliability, consistency, portability, naturalness,

and users’ subjective satisfaction. Human–computer in-

terface (or simply ‘‘user interface’’) is a software applica-

tion, a system that realizes human–computer interaction.

▶ Biometric System Ergonomic Design

▶ Facial Expression Recognition

Human-Interpretable Fingerprint

Classes

Fingerprints are grouped based on some visua l char-

acteristics of ﬁngerprint images determined by human

experts. Such groups are called human-interpretable

ﬁngerprint classes. An excellent example of human-

interpretable ﬁngerprint classes is the well-known

Galton–Henry classiﬁcation scheme proposed by Sir

Francis Galton and Edward Henry. The ﬁve most com-

mon Galton–Henry classes are called arch, tented arch,

left loop, right loop, and whorl, which are easy to be

understood even by ordinary people.

▶ Fingerprint Classification

Hypothesis Test

Hypothesis testing refers to the process of using statis-

tical analysis to determine if the observed differences

between two or more samples are due to random

chance (as stated in the background hypothesis) or

true differences in the samples (as stated in the target

hypothesis).

▶ Universal Background Models

Hypothesis Test

713

ICP Algorithm

Iterative Closest Point (ICP) algorithm developed by

Besl and Mckay, is a well-known method to align 3D

shapes. However ICP requires that ever y point in one

set have a corresponding point on the other set. This

cannot be guaranteed in practice. As a result modiﬁed

ICP algorithms exist in the literature.

▶ Ear Biometrics, 3D

ID Photograph

▶ Photography for Face Image Data

Identification

Biometric identiﬁcation is a process that ranks the

biometric references in the enrolment database in

order of decreasing similarit y against a recognition

biometric sample and then makes a decision, based

on the similarity scores, about the identity w.r.t. to

the references.

▶ Verifica tion/Identification/Authentication/Recogni-

tion: The Terminology

Identity Level in the Speech Signal

Speech production is an extremely complex process,

whose result depends on many variables at different

levels, including sociolinguistic factors (e.g., level of

education, linguistic context, and dialectal differences)

and morphological issues (e.g., vocal tract length

and shape or the dynamic conﬁguration of the articu-

latory organs). These multiple inﬂuences will be

simultaneously present in each speech act and some

or all of them w ill contain speciﬁcities of the speaker.

Hence, it is needed to clarify and clearly distinguish

the different levels and sources of speaker infor-

mation that should be extracted to model speaker

individualities.

▶ Speaker Features

Identity Theft Reduction

▶ Fraud Reduction, Overview

Illumination

Ambient light sources may affect the appearance of a

biometric image (such as face, ﬁngerprint, or iris). The

intensity and direction of these light sources can im-

pact the performance of image-based biometric recog-

nition algorithms.

▶ Biometrics, Overv iew

▶ Face Recognition, Near-Infrared

▶ Face Tracking

▶ Illumination Compensation

▶ Photography for Face Image Data

2009 Springer Science+Business Media, LLC

Illumination Compensation

XUDONG XIE

,KIN-MAN LAM

,QIONGHAI DAI

Automation Department, Tsinghua University,

Beijing, China

Department of Electronic and Information

Engineering, The Hong Kong Polytechnic University,

Hong Kong, China

Synonyms

Lighting compensation; Illumination normalization

Definition

Due to difﬁculty in controlling the lighting conditions

in practical applications, variable illumination is one

of the most challenging tasks in face recognition. Prior

to face recognition, illumination compensation has to

be performed, whereby the uneven illumination of

human faces is compensated and face images in normal

lighting conditions are reconstructed. The recon-

structed face images are then used for classiﬁcation.

An illumination compensation scheme includes the

following modules: lighting category evaluation, shape

normalization, and lighting compensation.

Introduction

Human face recognition, one of the most successful

applications of image analysis and understanding, has

received signiﬁcant attention in the last decade. Howev-

er, due to difﬁculty in controlling the lighting conditions

in practical applications, variable illumination is one of

the most daunting challenges in face recognition. As

stated by Adini et al. [1], ‘‘The variations between the

images of the same face due to illumination and view-

ing direction are almost always larger than image var-

iations due to change in face identity’’. Most of the

available methods for face recognition, such as the

Principal Component Analysis (PCA) [2], and the In-

dependent Component Analysis (ICA) [3], encounter

difﬁculties under varying lighting conditions. Some

images under varying illuminations presented Fig. 1,

all images are from the Yal eB databa se [4]. It can be

seen that although all these images are of the same per-

son, due to the effect of uneven lighting, they look quite

different. Therefore, when the images are under varying

illumination, illumination compensation should be

performed before face recognition. Here, ‘‘illumination

compensation’ ’ means compensation for the uneven illu-

minations on human faces and reconstruction of face

images in normal lighting conditions, then the recon-

structed face images are used for classiﬁcation.

Some algorithms have been proposed to reduce the

effect of uneven lighting from an image processing

point of view. Histogram equaliz ation (HE) is a com-

monly used method to convert an image so that it has a

uniform histogram, which is considered to produce an

‘‘optimal’’ overall contrast in the image. However, after

being processed by HE, the lighting condition of

an image under uneven illumination may sometimes

become even more uneven. Adaptive histogram equal-

ization (AHE) [5] computes the histogram of a local

image region centered at a given pixel to determine the

mapped value for that pixel leading to local contrast

enhancement. However, the enhancement often leads

to noise ampliﬁcation in ‘‘ﬂat’’ regions, and ‘‘ring’’

artifacts at strong edges. In addition, this technique is

computationally intensive. Zhu et al. [6] proposed an

illumination correction method, which uses an

afﬁne transformation lighting model based on a local

estimation of the background and the illumi nation

gain. However, the method is useful only when

the images are under slowly varying illumination. The

Illumination Compensation. Figure 1 Samples of cropped faces under varying illuminations. The azimuth angles of the

lighting of images from left to right column are: 0



,20



,35



,70



, 50



and 70



, respectively. The corresponding

elevation angles are: 20



,90



, 40



,65



, 35



, 40



and 45



, respectively.

716

Illumination Compensation

Block-based Histogram Equalization (BHE) method [7]

divides an image into a number of small blocks, and

histogram equalization is performed within each of

the image blocks. The BHE is simple, and the compu-

tation required is for less than that for AHE. However,

as in the case of the AHE, noises are also enhanced

after being processed by the BHE. The main idea

of Local Normalization (LN) [8] is to split the face

region into a set of triangular facets, and then the

intensity values within each facet are normalized to be

of zero mean and unit variance. This method is very

fast, but it is also sensitive to variations caused by

local shape distortions, such as expression variations.

On the basis of a model-based method, [7] the general

procedure for illumination compensation can be

understood.

Human Face Model and

Lighting Model

A face image is assumed to be a ▶ Lambertian surface,

which can be described by the product of the albedo

and the cosine angle between the point light source and

the surface normal as follows:

Ix; yðÞ¼r x; yðÞn x; yðÞs; ð1Þ

where I(x, y) is the intensity value observed of the

pixel at (x, y) in the image, 0  r(x, y)  1 is the

corresponding albedo, n(x, y) is the surface normal

direction, s is the light source direction, and its magni-

tude is the light source intensity. Suppose I(x, y) and

(x, y) represent the pixel intensity values at (x, y)of

the image under normal lighting conditions and the

image under a certain kind of illumination, and s and

are the corresponding light source directions, then

the corresponding illumination ratio image [9] can be

given as follows:

ðx; yÞ¼I

ðx; yÞ=Iðx; yÞ

¼ðrðx; yÞnðx; yÞs

Þ=ðrðx; yÞnðx; yÞsÞ

¼ðnðx; yÞs

Þ=ðnðx; yÞsÞ

¼ Aðx; yÞ; ð2Þ

where A(x, y) is determined by the surface normal

direction n(x , y) and the kind of illumination con-

cerned. From (2), the following can be obtained:

x; yðÞ¼Ax; yðÞIx; yðÞ: ð3Þ

If the effect of additive noise at each point (x, y)is

considered, the illumination model in (3) can be ex-

tended to the following:

x; yðÞ¼Ax; yðÞIx; yðÞþBx; yðÞ; ð4Þ

where A(x, y) and B(x, y) denote the multiplicative

noise and the additive noise for the pixel (x, y),

respectively.

From (2) and (4), it can be seen that A(x, y) and

B(x, y) are only determined by the shape information

of a human face and the lighting condition, which

hints us that if the shapes of identities are normalized

to be the same, the values of A(x, y) and B(x, y) are

invariant for different persons under the same lighting

conditions. Therefore, after performing shape normal-

ization, the values of A(x, y) and B(x, y) can be com-

puted point by point for a certain illumination. Then

these values can be used to undertake illumination

compensation.

According to the illumination categories used

in the YaleB databas e [4], the lig hting conditions

are divided into 65 categories. Each of the categories

has different azimuth angles and elevation angles of

the lighting. The azimuth angles in the database vary

from 130



to þ130



, and the elevation angle ranges

from 40



to þ90



. If both the azimuth angle and the

elevation angle are equal to 0



, it can be said that the

subject is under normal illumination.

Shape Normalization and Lighting

Compensation

Suppose that the pixel-wise correspondence between an

input image and a reference face image is known, which

can be determined by facial feature detection. The

input image can be separated into texture and shape

using a 2D face shape model [10]. The shape of a face is

coded as the displacement ﬁeld from the reference

image, and the texture denotes an intensity map,

which is produced by mapping the original image on

to the reference image. All texture images have the

same shape as the reference image. According to the

authors, uneven illumination compensation is done on

the texture image in order to avoid disturbing the

shape information on the original image, i.e., in (1),

all identities have the same surface normal direction

distribution n. After illumination compensation, the

Illumination Compensation

717

compensated texture and the original shape are com-

bined to obtain the reconstructed image.

It is a challenge to ﬁnd the pixel-wise correspon-

dence between two pictures, especially when they

are under uneven lighting conditions. Here, the posi-

tion of some facial feature points, such as the eyebrows,

eyes, nos e, and mouth, are ﬁrst determined manually

(Fig. 2). The displacements of these key points between

a facial image and the reference image are subsequen-

tly computed. The reference shape is obtained from

the average 10 size-normalized and aligned images

from t he YaleB database. Using a triangle-b ased

cubic interpolation method [11], the input image

is mapped to the reference shape model. Afte r proces-

sing the mapped texture, it can be mapped

backwards from the reference shape to that of the

original shape.

Lighting compensation is performed based on the

mapped texture, which has a normal shape. From (4),

the following is obtained

Ix; yðÞ¼

x; yðÞBx; yðÞ

Ax; yðÞ

: ð5Þ

In order to avoid overﬂowing, all the intensity values

of I(x, y) are restricted to the range of [0, 255], so (5)

can be rewritten as follows.

Iðx; yÞ¼

0; Iðx; yÞ < 0;

255; Iðx; yÞ > 255;

ðx; yÞBðx; yÞ

Aðx; yÞ

; otherwise:

ð6Þ

Therefore, for an input image, if the illumination cate-

gory is determined and the corresponding values of

A(x, y) and B(x, y) are precomputed, (6) can be used

to calculate the reconstructed image, which is under

normal lig hting conditions.

Besides the effect of illumination on appearance, face

images of distinct subjects actually look quite different.

This is because the appearance of a human face is also

dependent on other factors, such as gender, race, and

makeup. In order to accurately estimate the light source

category, the distinctive elements of a person’s appear-

ance have to be eliminated to the extent possible while

keeping the illumination information unchanged. Then,

the illumination map is used to determine the illumina-

tion category. An image processed by the BHE [7]is

considered as a reference image. The BHE processed

image is then compared to the same image processed

by the HE to obtain a pixel-wise difference between the

two images. This difference image, which is called an

illumination map, reﬂects the effect of the light source

on different parts of the face image, and can be used to

estimate the illumination category.

The determination of the illumination category

is done by (LDA)

▶ Linear Discriminant Analysis [12].

The training images are divided into 65 different

categories on the basis of their lighting conditions,

and each category includes nine images that are under

the same lighting condition and that belong to differ-

ent people in the YaleB database.

Based on the training images in the YaleB database,

the optimal values for A

(x, y) and B

(x, y) for each

illumination category can be estimate d by means of the

least-squared method. For each illumination category

i, suppose that the number of training samples equa ls

m, then (4) can be rewritten as follows:

ðx; yÞ

ðx; yÞ 1

ðx; yÞ



;

where i ¼ 1; ...; 65 and k ¼ 1; ...; m:

ð7Þ

Illumination Compensation. Figure 2 Facial feature

points that are used to build a pixel-wise correspondence.

718

Illumination Compensation

Let F

¼½I

ðx; yÞI

ðx; yÞ

; where T

represents the transpose, I

x; yðÞis the kth subject

under the ith lighting category in the training set, and

F ¼

ðx; yÞ 1

;

where I

(x, y) represents the face of the kth subject in

the training set under normal lighti ng conditions.

Then (7) can be written as follows:

¼ F

ðx; yÞ



; i ¼ 1; ...; 65: ð8Þ

As the images in the different row of F, i.e., I

(x, y), are

images of different people, they are independent of

each other. The least-squared solutio n to (8) can be

calculated as follows:

ðx; yÞ



¼ F



1

; i ¼ 1; ...; 65: ð9Þ

Using (9), the optimal value of A

(x, y) and B

(x, y) for

the ith lighting category can be computed, and A

(x, y)

and B

(x, y) are called A-map and B-map, respectively.

(For more detailed descriptions, refer to Xie and

Lam [7]).

Experimental Results

The training of the illumination compensation algo-

rithm is based on the YaleB database. Therefore, the

performance of the algorithm can be evaluated by

utilizing other databases such as the Yale face database,

[13] the YaleB face database, and the AR face data-

base [14]. For each database, only images with an

upright frontal view and a neutral expression are se-

lected. The original images in the databases are shown

in the ﬁrst row of Figs. 3–5, images processed by the

HE in the second row, those processed by the BHE in

the third row, and images processed by the introduced

algorithm in the fourth row.

After illumination compensation, the recon-

structed images are used for face recognition. The

PCA is used to measure the recognition rates after

processing the images by different illumination com-

pensation techniques. In each database, one image for

each subject with normal illumination is selected as a

Illumination Compensation. Figure 3 Some experimental results based on the YaleB database. (a) original images,

(b) images processed by the HE, (c) images processed by the BHE, (d) images processed by the introduced algorithm.

Illumination Compensation

719