LIGHTING-VARIABLE ADABOOST BASED-ON SYSTEM

FOR ROBUST FACE DETECTION

R. Wood and J. I. Olszewska

School of Computing and Engineering, University of Huddersﬁeld, Queensgate, Huddersﬁeld, HD1 3DH, U.K.

Keywords:

Face detection, AdaBoost, Global intensity average value, Illumination variation, Lighting measure.

Abstract:

In order to detect faces in pictures presenting difﬁcult real-world conditions such as dark background or back-

lighting, we propose a new method which is robust to varying illuminations and which automatically adapts

itself to these lighting changes. The proposed face detection technique is based on an efﬁcient AdaBoost

super-classiﬁer and relies on multiple features, namely, the global intensity average value and the local inten-

sity variations. Based on tests carried out on standards datasets, our system successfully performs in indoor as

well as outdoor situations with different lighting levels.

1 INTRODUCTION

Face detection is a very important and popular ﬁeld

of research in computer vision as it is usually the ﬁrst

step of applications such as automatic human recog-

nition, facial expression analysis, identity certiﬁca-

tion, trafﬁc monitoring or advanced digital photogra-

phy. For that, many techniques have been developed

based on different features of a human face such as

the color of the skin (Gundimada et al., 2004), its tex-

ture (Ahonen et al., 2006) or both (Woodward et al.,

2010). Some methods rely on motion detection, since

the eye are blinking or the lips are moving (Crow-

ley, 1997). Other approaches are based on active

contour (Olszewska et al., 2008) techniques to delin-

eate the shape of the face (Yokoyama et al., 1998),

the mouth (Li et al., 2006) or the hairs (Julian et al.,

2010). Some works consider as facial features char-

acteristic parts of the human face such as eyes (Lin

et al., 2008) or ears (Hurley et al., 2008). The feature

classiﬁcation is usually done by means of neural net-

works (Rowley et al., 1998), Hausdorff distance mea-

sure (Guo et al., 2003), AdaBoost algorithm (Viola

and Jones, 2004) or support vector machine method

(Heisele et al., 2007).

One of the main issues of the face detection tech-

niques is their sensitivity to the lighting variations

of the background and/or foreground. For example,

methods based on skin color do not perform efﬁ-

ciently in case of foreground darkness in the picture

(Zhao et al., 2003). Despite some recent works such

as (Sun, 2010) or (Huang et al., 2011) which attempt

to improve the automatic process of face detection

in still pictures, face detection robust towards illumi-

nation changes is still a challenging task (Beveridge

et al., 2010).

In this work, we have tackled with face detection

in variable illumination conditions. For our purposes,

we have developed an innovative approach which au-

tomatically adapts itself to these lighting changes in

order to increase the robustness of the detection sys-

tem.

The contribution of this paper is two-fold. Indeed,

we present a new super-classiﬁer which is based on

two strong classiﬁers and furthermore, which allows

the combined use of two variables. Hence, on one

hand, the global image intensity is calculated and, on

the other hand, local features such as Haar-like ones

are extracted. The proposed super-classiﬁer relying

on both these global and local image features leads to

the efﬁcient detection of faces in any types of color

images, especially in difﬁcult situations with back-

ground or foreground presenting low levels of light-

ing.

The paper is structured as follows. In Section 2,

we present our Adaboost-based face detection method

which embeds two modalities, namely, the global im-

age intensity and the Haar-like features measuring lo-

cal changes in intensity. The resulting system that

aims to automatically detect faces in images with dif-

ferent illumination conditions has been successfully

tested on real-world standard images as reported and

discussed in Section 3. Conclusions are drawn up in

Section 4.

494

Wood R. and I. Olszewska J..

LIGHTING-VARIABLE ADABOOST BASED-ON SYSTEM FOR ROBUST FACE DETECTION.

DOI: 10.5220/0003888304940497

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (MPBS-2012), pages 494-497

ISBN: 978-989-8425-89-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

(a) (b)

Figure 1: Our system’s performance in the case of face detection in images with (a) artiﬁcial lighting or (b) day light.

2 LIGHTING-ADAPTABLE FACE

DETECTION

Our face-detection system is based on two modali-

ties that are the global image intensity and the local

Haar-like features, described in Sections 2.1 and 2.2,

respectively. The proposed AdaBoost-based method

which combines these two information is explained

in Section 2.3.

2.1 Global Image Intensity Average

Value

Let us consider a color image I(x, y) with M and N,

its height and width, respectively. The conversion of

the image I(x, y) to the gray one I

(x, y) according to

the ITU-R BT. 601 norm (Kawato and Ohya, 2000) is

as follows

(x, y) = 0.299R(x, y)+0.587G(x, y)+0.114B(x, y),

(1)

where R, G, and B are the red, green and blue

channels, respectively, of the initial color image

I(x, y).

Based on the gray image I

(x, y), we deﬁne the

average value I

AV G

of the global image intensity as

AV G

M N

∑

x=1

∑

y=1

(x, y).

(2)

Hence, in opposition to the other lighting compen-

sation methods, e.g. mentioned in (Gundimada et al.,

2004), (Huang et al., 2011), which directly change the

pixel values of the original image, we compute a sin-

gle average value I

AV G

of the global intensity of the

gray image and we use it as a pivot as explained in

Section 2.3. In fact, our approach shows better face

detection rates as demonstrated in Section 3.

2.2 Haar-like Features

Local Haar-like features f (Viola and Jones, 2004)

encode the existence of oriented contrasts between re-

gions in the processed image. They are computed by

subtracting the sum of all the pixels of a subregion of

the local feature from the sum of the remaining region

of the local feature using the integral image represen-

tation II(x, y) which is deﬁned as follows

II(x, y) = II(i − 1, j) + II(i, j − 1) +I(i, j), (3)

where I(i, j) is the pixel value of the original im-

age at the position (i, j).

2.3 Lighting-variable AdaBoost

Detection (LVAD) System

At ﬁrst, our LVAD detection system requires a train-

ing phase during which it builds strong classiﬁers

based on cascades of weak classiﬁers.

In particular, to form a T -stage cascade, T weak

classiﬁers are selected using the AdaBoost algorithm

(Viola and Jones, 2004). In fact at a t stage of this cas-

cade, a sub-window u of an image from the training

set is computed by eq. (3) and it is passed to the corre-

sponding t

weak classiﬁer. If the region is classiﬁed

as a non-face, the sub-window is rejected. If not, it is

passed to the t +1 stage, and so forth. Consequently,

more stages the cascade owns, more selective it is,

i.e. less false positive detections occur. However, that

could lead to the increase in the number of false neg-

ative detection.

In order to select at each t level (with 1 < t < T)

the best weak classiﬁer, an optimum threshold θ

computed by minimizing the classiﬁcation error due

to the selection of a particular Haar-like feature value

LIGHTING-VARIABLE ADABOOST BASED-ON SYSTEM FOR ROBUST FACE DETECTION

495

(a) (b)

Figure 2: Our system’s performance in the case of face detection in images with (a) dark background or (b) backlighting.

(u). The resulting weak classiﬁer k

is thus obtained

as follows

(u, f

, p

, θ

) =

(

1 if p

(u) < p

0 otherwise,

(4)

where p

is the polarity indicating the direction of

the inequality.

Then, a strong classiﬁer K

(u) is constructed by

taking a weighted combination of the selected weak

classiﬁers k

according to

(u) =

(

1 if

∑

t=1

(u) ≥

∑

t=1

0 otherwise,

(5)

where

∑

t=1

is the AdaBoost threshold and α

is a voting coefﬁcient computed based on the classi-

ﬁcation error in each stage t of the T stages of the

cascade (Viola and Jones, 2004).

Next, we introduce the lighting-adaptable strong

super-classiﬁer K (u) deﬁned as

K (u) =

(

(u) if I

AV G

> I

(u) if I

AV G

≤ I

(6)

where I

is the global image intensity threshold

and where D and L (with D ≥ L) are the numbers of

the weak classiﬁers for dark and light images, respec-

tively.

In this way, the proposed LVAD system allows to

automatically select the number of stages of the Ad-

aBoost cascade accordingly to the lighting conditions

expressed in eq. (6) by I

AV G

During the testing phase, the LVAD trained system

is applied to detect faces/non-faces in a test image set

which does not contain any of the images of the train-

ing set. Haar-like features are thus extract from these

test images according to eq. (3) and classiﬁed using

the lighting-adaptable strong super-classiﬁer K (u) as

deﬁned in eq. (6). The resulting face detection shows

excellent performance as discussed in Section 3.

3 RESULTS AND DISCUSSION

We have tested our system on standard datasets such

as (Fei-Fei et al., 2003). We have ﬁrst trained our clas-

siﬁer on four positive and four negative images with

two different numbers of weak classiﬁers. Then, our

system was applied to detect faces in all the images of

the database.

Some examples of the performance of our ap-

proach for face detection in indoor and outdoor scenes

with dim light as well as in case of darkness are pre-

sented in Figs. 1 and 2, respectively.

To quantitatively assess the obtained results, the

detection rate (Huang et al., 2011) is deﬁned as

detection rate =

T P

T P +FN

, (7)

with TP, true positive and FN, false negative.

Table 1: Face detection rates.

(Viola and Jones, 2004) (Huang et al., 2011) LVAD

91% 94.4% 95%

The results of these experiments are reported in

Table 1. The overall detection rate of our LVAD sys-

tem is 95% and our approach characterized by an ad-

justable number of weak classiﬁers is well robust for

the different lighting levels present in the dataset im-

ages. Moreover, our method outperforms the state-

of-the art algorithm of (Viola and Jones, 2004) which

owns a ﬁxed number of weak classiﬁers and the recent

work of (Huang et al., 2011) for varying illumination.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

496

4 CONCLUSIONS

In this work, we have proposed an efﬁcient and ro-

bust face detection system that uses our strong super-

classiﬁer based on Adaboost cascades with an adapt-

able number of weak classiﬁers which is depending

on the illumination conditions of the captured image.

In the future, we aim to replace in our system

the global-lighting average value which computation

is currently based on the processed image with a di-

rect real-world lighting measure recorded by sensitive

sensors.

REFERENCES

Ahonen, T., Hadid, A., and Pietikainen, M. (2006). Face

descritpion with local binary patterns: application to

face recognition. IEEE Transactions on Pattern Anal-

ysis and Machine Intelligence, 28(12):2037–2041.

Beveridge, J. R., Bolme, D. S., Draper, B. A., Givens, G. H.,

Liu, Y. M., and Phillips, P. J. (2010). Quantifying

how lighting and focus affect face recognition per-

formance. In IEEE Computer Society Conference on

Computer Vision and Pattern Recognition Workshops,

pages 74–81.

Crowley, J. L. (1997). Vision for man machine interaction.

Robotics and Autonomous Systems, 19(3-4):347–359.

Fei-Fei, L., Andreetto, M., and Ranzato,

M. A. (2003). The Caltech - 101 Ob-

ject Categories dataset. Available online:

http://www.vision.caltech.edu/feifeili/Datasets.htm.

Gundimada, S., Tao, L., and Asari, V. (2004). Face detec-

tion technique based on intensity and skin color dis-

tribution. In IEEE International Conference in Image

Processing, volume 2, pages 1413–1416.

Guo, B., Lam, K.-M., Lin, K.-H., and Siu, W.-C. (2003).

Human face recognition based on spatially weighted

Hausdorff distance. Pattern Recognition Letters, 24(1-

3):499–507.

Heisele, B., Serre, T., and Poggio, T. (2007). A component-

based framework for face detection and identiﬁcation.

International Journal of Computer Vision, 74(2):167–

181.

Huang, D.-Y., Lin, C.-J., and Hu, W.-C. (2011). Learning-

based face detection by adaptive switching of skin

color models and AdaBoost under varying illumina-

tion. Journal of Information Hiding and Multimedia

Signal Processing, 2(3):2073–4212.

Hurley, D. J., Harbab-Zavar, B., and Nixon, M. S. (2008).

Handbook of Biometrics, chapter The ear as a biomet-

ric, pages 131–150. Springer-Verlag.

Julian, P., Dehais, C., Lauze, F., Charvillat, V., Bartoli, A.,

and Choukroun, A. (2010). Automatic hair detection

in the wild. In IEEE International Conference on Pat-

tern Recognition, pages 4617–4620.

Kawato, S. and Ohya, J. (2000). Real-time detection of nod-

ding and head-shaking by directly detecting and track-

ing the ”Between-Eye”. In IEEE International Con-

ference on Automatic Face and Gesture Recognition,

pages 40–45.

Li, Y., Lai, J. H., and Yuen, P. C. (2006). Multi-template

ASM method for feature points detection of facial im-

age with diverse expressions. In IEEE International

Conference on Automatic Face and Gesture Recogni-

tion, pages 435–440.

Lin, K., Huang, J., Chen, J., and Zhou, C. (2008). Real-time

eye detection in video streams. In IEEE International

Conference on Natural Computation, volume 6, pages

193–197.

Olszewska, J. I., DeVleeschouwer, C., and Macq, B. (2008).

Multi-feature vector ﬂow for active contour track-

ing. In IEEE International Conference on Acoustics,

Speech and Signal Processing, pages 721–724.

Rowley, H. A., Baluja, S., and Kanade, T. (1998). Neural

network-based face detection. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 20(1):23–

38.

Sun, H.-M. (2010). Skin detection for single images using

dynamic skin color modeling. Pattern Recognition,

43(4):1413–1420.

Viola, P. and Jones, M. J. (2004). Robust real-time face

detection. International Journal of Computer Vision,

57(2):137–154.

Woodward, D. L., Pundlik, S. J., Lyle, J. R., and Miller,

P. E. (2010). Periocular region appearance cues for

biometric identiﬁcation. In IEEE Computer Society

Conference on Computer Vision and Pattern Recogni-

tion Workshops, pages 162–169.

Yokoyama, T., Yagi, Y., and Yachida, M. (1998). Active

contour model for extracting human faces. In IEEE

International Conference on Pattern Recognition, vol-

ume 1, pages 673–676.

Zhao, W., Chellappa, R., Phillips, P. J., and Rosenfeld, A.

(2003). Face recogntion: A literature survey. ACM

Computing Surveys, 35(4):399–458.

LIGHTING-VARIABLE ADABOOST BASED-ON SYSTEM FOR ROBUST FACE DETECTION

497