CLASSIFIER SELECTION FOR FACE RECOGNITION

ALGORITHM BASED ON ACTIVE SHAPE MODEL

Andrzej Florek and Maciej Król

Institute of Control and Information Engineering, Automatic Control and Robotics Division

Poznań University of Technology, 60-965 Poznań, str. Piotrowo 3A, Poland

Keywords: Face classification, Active Shape Model, Support Vector Machines, Fisher Linear Discriminant Analysis.

Abstract: In this paper, experimental results from the face contour classification tests are shown. The presented

approach is dedicated to a face recognition algorithm based on the Active Shape Model. The results were

obtained from experiments carried out on the set of 2700 images taken from 100 persons. Manually fitted

contours (194 samples for eight components of one face contour) were classified after feature space

decomposition carried out by the Linear Discriminant Analysis or by the Support Vector Machines

algorithms.

1 INTRODUCTION

The presented algorithm for a face classification is

based on the Active Shape Model method (ASM)

(Cootes, 2001), which is a modification of the

Active Contour Model method (Kass, 1988), i.e.

a snake-based approach to extracting face contours

from an image. ASM is based on a shape notation,

which is defined as an ordered set of points and it is

a two-stage algorithm. First, a Point Distribution

Model (PDM) is produced to be used for the

validation of a contour shape. Next, a Local Gray

Level Model (LGLM) is generated for interactive

fitting the contour points to the local image context.

To apply the ASM, an initial contour and its

preliminary location have to be known. This method

is still in progress. Modifications consist of initial

contour choice and the new fitting methods (Zuo,

2004), (Zhao, 2004).

To obtain PDM and LGLM, the desirable

contour localisation on a real image has to be

known. Thus, placing contours onto images chosen

to create a learning set has to be performed. It may

be done manually or semi-automatically. In

presented paper, manually placed contours were

used for testing classifiers. Two methods for the face

contour classification to any class were examined.

The first method was the Nearest Neighbourhood

Classifier (NNC) in reduced Fisher feature subspace

(Linear Discriminant Analysis – LDA) with

Euclidean distance. The second method was the

Support Vector Machines (SVM) with a voting

system or with a criterion based on a maximal

distance from separating classes hyperspaces. A set

of 2700 contours was used to create the learning and

the validation sets and to account classification

accuracy.

2 CONTOURS

The shape in ASM method is represented as an

ordered set of control points placed on contours

describing face elements and it is given by the

following vector

x = ( x

, y

, x

, y

,..., x

, y

)

, (1)

Table 1: Face contours.

Contour Number of points

Face outline 41

Mouth outer 28

Mouth inner 28

Right eyelid 20

Left eyelid 20

Right eyebrow 20

Left eyebrow 20

Nose outline 17

TOTAL 194

where x

and y

are coordinates of shape control

points, expressed in common coordinate frame, for

all shapes in a given set.

276

Florek A. and Król M. (2008).

CLASSIFIER SELECTION FOR FACE RECOGNITION ALGORITHM BASED ON ACTIVE SHAPE MODEL.

In Proceedings of the Third International Conference on Computer Vision Theory and Applications, pages 276-281

DOI: 10.5220/0001077802760281

 SciTePress

In the considered case in the paper, for eight

component face contours of interest, n = 194 such

points have been determined (Fig. 1a and Tab. 1)

and this implies 388 dimensional shape space.

2.1 Extracting and Calculation of

Contours

The stages of applied procedure to obtain normalised

contours are presented in Fig. 1. Images used to

contour extracting are presented in Fig. 2. First, two

landmarks are positioned on the face image, in the

external eye corners. Next, the initial contour

(template) is placed on the image, according to the

landmark positions (Fig. 1a). The landmarks

determine a face pose and an image scale. In the

next step, the contour is manually drawing (fitting)

on place, which seems to the operator as the best

localisation for the contour point positions (Fig. 1b).

Subsequently, the derived contours (Fig. 1c) are

normalised. Scale coefficient results from the

calculated coordinates of eye centres (pupils). This

is connected to applied active shape procedure,

where the initial contour is generally placed

according to expected pupils positions. Pupil

coordinates are calculated from coordinates of

contour samples located in the eyelid corners. X-axis

is determined by pupil coordinates; points (-1,0),

(1,0) are located on right and left pupil, respectively.

The symmetrical of this section determines Y-axis

and the middle of coordinate system. Next, the

contour points are projected on a normalised

coordinate system. The normalised contour has to be

uniformly sampled (manually extracted contours

have nonuniform distances between the adjoining

points). The normalised and uniformly sampled

contours are presented in Fig. 1d. During the

normalisation procedure, points are ordered in

a defined sequence, according to the feature vector

definition (1). In the presented approach, the height

standardizations of face and nose outlines have not

been applied (Fig. 3).

3 EXPERIMENT

In order to select classifiers for ASM method, an

experiment consisting of examining a set of face

images was undertaken. Color images of 2048 ×

1536 pixels were used. For 100 persons (N = 100

classes) the following images were taken:

A – sequence of 30 frames for horizontal head

rotation from the right to the left half-profile;

B – sequence of 20 frames for vertical face rotation

from slightly risen to hanged down head

position;

C – 10 frames for different head position and

limited face mimicry.

The contours were prepared by over a dozen

persons. A person chosen to work on C-frames has

not seen the contours resulting from A and B frames.

The contours were positioned on 11 internal images

from A-frames, on 11 internal images from B-frames

and on 5 images chosen from C-frames (Fig. 2). In

presented experiment 2700 contours were used.

3.1 Set Definitions

The normalised contours were divided into the

following sets:

 LS22 – learning set, 2200 contours, 22 for each

class from A- and B-frames;

 LS11A – learning set, 1100 contours, 11 for

each class, even subset of LS22;

 LS11B – learning set, 1100 contours, 11 for

each class, odd subset of LS22;

 LS06 – learning set, 600 contours, 6 for each

class, subset of LS22 with face poses nearest

to en face position;

 VS05 – validation set, 500 contours, 5 for each

person from C-frames.

LS11A and LS11B sets were used as learning or

testing sets alternatively.

3.2 Classifiers

Two classifying methods were tested. The first

classifier was Nearest Neighbourhood Classifier in

reduced shape subspace derived from LDA. As

a metric, Euclidean distance to a model of class in

99-dimensional subspace was used. The second

method was taken as the SVM method with kernel

such as Radial Basis Function. The classification of

x sample from testing or learning sets was based on

a voting procedure. In presented approach, a total

number of votes is equal to N (N - 1)/2, where N is

the number of classes. The maximal number of votes

to one class is equal to (N - 1) and in our experiment

it is only 2% of total number of votes. The voting

decision depends only on the sign of discrimination

function for x sample coordinates. In the case of

a pair of “very similar classes”, only one vote from

(N - 1) decisions can decide. In the presented

CLASSIFIER SELECTION FOR FACE RECOGNITION ALGORITHM BASED ON ACTIVE SHAPE MODEL

277

Figure 1: Face contours: a) initial contour and its position on the image, b) images with manually fitted contours,

c) extracted contours, d) normalized contours.

VISAPP 2008 - International Conference on Computer Vision Theory and Applications

278

a) b) c)

Figure 2: Images from learning set LS22 (a, b) and validation set VS05 (c): a) boundary frames for sequences symmetrical

to “en face” position, b) boundary frames for nonsymmetrical sequences, c) images from validation set VS05.

example, 70% first succeeding votes for one sample

were: 99, 98 and 97. That is why, one other classifier

for SVM was proposed - the maximal total distance

from all demarcated hyperspaces. Total distance

(margin) td

(x) for C

class is calculated as

(x) = ∑ H

(x) ,

= 1

≠

(2)

where H

(x) is a decision function for the pair of

classes (C

, C

) and the value is positive if x has been

classified to class C

and negative if, it has been

classified to Cj ( H

(x) = 0 is the equation of

boundary hyperspace). The vector x is classified to

the class with maximal td

(x) value. The denoting

values of elements in the decision function matrix

between C

and Cj classes by decision(i, j), voting

algorithm is, as follow:

CLASSIFIER SELECTION FOR FACE RECOGNITION ALGORITHM BASED ON ACTIVE SHAPE MODEL

279

Figure 3: Normalised contours of three people from learning set LS22 (a) and validation set VS05 (b).

IF decision(i,j) > 0

vote(i) += abs(decision(i,j))

vote(j) -= abs(decision(i,j))

ELSE

vote(i) -= abs(decision(i,j))

vote(j) += abs(decision(i,j))

END

The vector x is classified to the class indicated by

the value of

argmax(vote) function.

3.3 Results

As accuracy measure of classifiers, the coefficient

TP/(TP+FP) in percent was chosen, where TP and

FP are True Positive and False Positive numbers of

final classifications. Classification was performed

using 3 methods, named as:

 LDA – Fisher Discriminant Analysis and

Euclidean distance in reduced feature space;

 SVM-v – SVM and voting system 1-1;

 SVM-d – SVM and maximal distance

criterion.

Results are presented in Tab. 2 (for learning and

testing sets) and in Tab. 3 (for validation set).

Table 2: Classification accuracy for learning and testing

sets (in %).

Learning set – testing set LDA SVM-v SVM-d

LS11A – LS11B 99,9 99,3 95,4

LS11B – LS11A 96,5 99,7 97,3

Table 3: Classification accuracy for validation set (in %).

Learning set LDA SVM-v SVM-d

LS22 94,4 78,6 57,0

LS11A 88,2 76,4 58,0

LS11B 68,8 76,6 57,6

LS06 52,8 74,8 53,0

4 SUMMARY

Results in Tab. 2 confirm good propriety of

classifiers, however this is the situation where the

learning and testing sets are nearly regular subsets of

larger learning set LS22. Thus, it is possible to apply

these algorithms to an automatic recognition system

when people want to be recognized. Results in

Tab. 3 are based on tests over the validation set

VS05. Contours belonging to VS05 were manually

posed on others images, not so regular as those in

the learning set LS22 and they were delivered by

other operators. The accuracy in LDA decreases for

smaller learning sets. Lower accuracy for LS11B set

compared with LS11A set possibly results from

different number of images selected from A- and B-

frames. The learning set LS11A has six contours

from A-frames and five contours from B-frames and

LS11B set inversely. The validation set VS05

consisted of frames more similar to A-frames.

Results for SVM methods are rather the same and

whole inferior to LDA. Only for little learning set

VISAPP 2008 - International Conference on Computer Vision Theory and Applications

280

LS06, SVM methods is significant better that LDA

method (more than 40%). The proposed SVM-d

method did not improve classification results. This

may suggest that LDA method is more resistant to

diversity of validation set because the space

transformation function is found in order to

maximize the ratio of between-class variance to

within-class variance. Classifiers based on SVM

transformations, require distinctly representative

learning set. SVM classification is most laborious,

operates in dimensionally higher space and requires

larger voting number than LDA. Presented

experiment shows that even for large but

homogeneous learning set (with relatively small

variance) and various, heterogeneous validation set

(practically normal situation in a visual inspection

system) face classification algorithm based on linear

discriminant analysis seems to be still advisable.

It is desirable to examine influence of other

contour normalisation procedures and to reduplicate

presented experiment, taking into consideration the

contours, automatically calculated by the trained

ASM algorithm. It would be interesting to analyse

how the number of classes N influences the accuracy

of LDA and SVM algorithms.

In presented experiment, the standardisation of

face outline and

nose outline heights has not been

(e.g. to pupils line position). Other normalisation

procedures, application of initial contour determined

by calculated face position (Ge, 2006) and identified

face gestures (de la Torre, 2007) will be verified in

the feature research.

REFERENCES

Cootes, T. and Taylor, C., (2001). Statistical models of

appearance for computer vision, Technical report,

University of Manchester, Wolfson Image Analysis

Unit, Imaging Science and Biomedical Engineering.

Ge, X., Yang, J., Zheng, Z., Li, F., (2006). Multi-view

based face chin contour extraction, Engineering

Applications of Artificial Intelligence, vol.19, pp. 545-

555.

Kass, M., Witkin, A., Terzopoulos, D., (1988). Snakes:

Active contour models, International Journal of

Computer Vision, 1 (4), pp. 321-331.

de la Torre, F., Campoy, J., Coha, J. F., Kanade, T.,

(2007). Simultaneous registration and clustering for

temporal segmentation of facial gestures from video,

Proceedings of the Second International Conference

on Computer Vision Theory and Applications

VISAPP 2007, vol. 2 IU/MTSV, pp. 110-115.

Zhao, M., Li, S. Z., Chen, Ch., Bu, J., (2004). Shape

evaluation for weighted active shape models,

Proceedings of the Asian Conference on Computer

Vision, pp 1074-1079.

Zuo, F. and de With, P. H N., (2004). Fast facial feature

extraction using a deformable shape model with Haar-

wavelet based local texture attributes, Proceedings of

ICIP’04, vol. 3, pp. 1425-1428.

CLASSIFIER SELECTION FOR FACE RECOGNITION ALGORITHM BASED ON ACTIVE SHAPE MODEL

281