Vol.1 ● No.

3 ● 2012 Scientific Research Journal of India 60

Face Exposure Technology

Thanigaivel.V*

Abstract:
The Face recognition is concerned with determining which part of an image contains
a face. If present, return the image location and content of each face. The automatic
system that analyzes the information contained in faces. While earlier works deal
primarily with standing front faces, several systems have been developed that are
able to detect faces reasonably truly plane or out-of-plane rotations in real time.
Even if a face exposure module is normally designed to deal with single images, its
performance can be improved if video capture.

INTRODUCTION

The technology has facilitated the foundation, faces need to be located and
development of real-time visualization registered first to facilitate further
modules that interact with humans. For processing. It is evident that face detection
biometric systems that use faces as non- plays an important and critical role for the
intrusive input modules, it is imperative to success of any face processing systems. The
locate faces in a picture before any face detection problem is testing as it needs
recognition algorithm can be applied. A to account for all possible look difference
vision based user interface should be able to caused by change in lights, facial features,
tell the attention focus of the user in order to occlusions. In addition, it has to detect faces
respond as a result. To detect facial features that appear at different technology, with in
truly for applications such as digital plane revolution. In spite of all these

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 61

difficulty, great progress has been made in detected faces are usually further processed
the last decade and many systems have to combine overlapped results and remove
shown inspiring real-time act. The recent false positives with heuristics1 or further
advances of these algorithms have also processing (e.g., edge exposure and
made major help in detecting other objects intensity variance). Numerous
such as humans, representations have been proposed for face
exposure, including pixel-based1, 3, 5, parts-
Face Exposure System based4, 6, 7
, local edge features8, 9, Haar
Most exposure systems carry out the task by wavelets4,10, and Haar-like features2, 11
.
extracting certain properties of a set of While earlier holistic representation
training images acquired at a fixed pose in schemes are able to detect faces1, 3, 5
, the
an off-line setting. To reduce the effects of recent systems with Haar-like features2, 12, 13
illumination change, these images are have demonstrated impressive empirical
processed with histogram equalization1, 3
results in detect faces under occlusion. A
Based on the extracted properties, these large and representative training set of face
systems typically scan through the entire images is essential for the success of
image at every possible location and scale learning-based face detector. From the set
in order to locate faces. The extracted of collected data, more positive examples
properties can be either manually coded or can be synthetically generated by perturbing;
learned from a set of data as adopted in the mirroring, rotating and scaling the original
recent systems that have demonstrated face images1, 3. On the other hand, it is
impressive results1, 2, 3, 4, 5. In order to detect relatively easier to collect negative
faces at different scale, the detection examples by randomly sampling images
process is usually repeated to a pyramid of without face images1, 3. As face exposure
images whose resolution is reduced by a can be mainly formulated as a pattern
1, 3
certain factor (1.2) from the original one . recognition problem, numerous algorithms
Such procedures may be expedited when have been proposed to learn their generic
other visual cues can be accurately templates (e.g., eigenface and statistical
incorporated (motion) as pre-processing distribution) or discriminate classifiers (e.g.,
5
steps to reduce the search space . As faces neural networks, Fisher linear discriminate,
are often detected across scale, the raw sparse network of Winnows, decision tree,

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 62

Bays classifiers, support vector machines, like features (at different position and scale)
and AdaBoost). Typically, a good face is very large (about 160,000). Contrary to
detection system needs to be trained with most of the prior algorithms that use one
several iterations. One common method to single strong classifier (e.g., neural
further improve the system is to bootstrap a networks and support vector machines),
trained face detector with test sets, and re- they used an ensemble of weak classifiers
train the system with the false positive as where each one is constructed by
well as negatives1. This process is repeated shareholding of one Haar-like feature. The
several times in order to further improve the weak classifiers are selected and weighted
performance of a face detector. A survey on using the AdaBoost algorithm14. As there is
these topics can be found in5, and the most large number of weak classifiers, they
recent advances are discussed in the next presented a method to rank these classifiers
section. into several cascades using a set of
optimization criteria. Within each stage, an
Recent technology ensemble of several weak classifiers is
The AdaBoost-based face detector by Viola trained using the AdaBoost algorithm. The
and Jones2 demonstrated that faces can be motivation behind the cascade of classifier
fairly reliably detect in real-time (i.e., more is that simple classifiers at early stage can
than 15 frames per second on 240 by filter out most negative examples efficiently,
320images with desktop computers) under and stronger classifiers at later stage are
partial occlusion. While Haar wavelets were only necessary to deal with instances that
used in10 for representing faces and look like faces. The final detector, a 38
pedestrians, they proposed the use of Haar- layer cascade of classifiers with 6,060 Haar-
like features which can be computed like features, demonstrated impressive real-
efficiently with integral image2. Figure 1 time performance with fairly high detection
shows four types of Haar-like features that and low false positive rates. Several
are used to encode the horizontal, vertical extensions to detect faces in multiple views
and diagonal intensity information of face with in-plane ration have since been
images at different position and scale. proposed12, 13, 15. An implementation of the
Given a sample image of 24 by 24 pixels, AdaBoost-based face detector2 can be found
the exhaustive set of parameterized Haar- in the Intel Open CV library. Despite the

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 63

excellent run-time performance of boosted selection algorithm used as a pre-computing

2
cascade classifier , the training time of such procedure, they reported that the training
a system is rather lengthy. In addition, the time of the classifier cascade with AdaBoost
classifier cascade is an example of is reduced by 50 to 100 times. For learning
degenerate decision tree with an unbalanced in each stage (or node) within the classifier
data set (i.e., a small set of positive cascade, they also exploited the asymmetry
examples and a huge set of negative ones). between positive and negative data using a
Numerous algorithms have been proposed linear classifier with the assumptions that
to address these issues and extended to they can be modeled with Gaussian
detect faces in multiple views. To handle distributions17. The merits and drawbacks of
the asymmetry between the positive and the proposed linear asymmetric classifier as
negative data sets, Viola and Jones well as the classic Fisher linear discriminate
proposed the asymmetric AdaBoost were also examined in their work. Recently,
algorithm16 which keeps most of the Pham and Champ proposed an online
weights on the positive examples. In 2, the algorithm that learns asymmetric boosted
AdaBoost algorithm is used to select a classifiers18 with significant gain in training
19
specified number of weak classifiers with time. In , an algorithm that aims to
lowest error rates for each cascade and the automatically determine the number of
process is repeated until a set of classifiers and stages for constructing a
optimization criteria (i.e., the number of boosted ensemble was proposed. While a
stages, the number of features of each stage, greedy optimization algorithm was
and the detection/false positive rates) is employed in 2, Brubaker et al. proposed an
satisfied. As each weak classifier is made of algorithm for determining the number of
one single Haar-like feature, the process weak classifiers and training each node
within each stage can be considered as a classifier of a cascade by selecting operating
feature selection problem. Instead of points within a receiver operator
20
repeating the feature selection process at characteristic (ROC) curve . The solved the
17
each stage, Wu et al. presented a greedy optimization problem using linear programs
algorithm for determining the set of features that maximize the detection rates while
for all stages first before training the satisfying the constraints of false positive
cascade classifier. With the greedy feature rates19. Although the original four types of

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 64

Haar-like features are sufficient to encode which consists of a series of detectors

upright frontal face images, other types of trained with positive images only, and the
features are essential to represent more energy-based method25 that simultaneously
complex patterns (e.g., faces in different detects faces and estimates their pose in real
11,12,13,15,21
pose) . Most systems take a divide- time.
and-conquer strategy and a face detector is
constructed for a fixed pose, thereby Quantifying Performance
covering a wide range of angles (e.g., yaws There are numerous metrics to gauge the
and pitch angles). A test image is either sent performance of face detection systems,
to all detectors for evaluation or to a ranging from detection frame rate, false
decision module with a coarse pose positive/negative rate, number of classifier,
estimator for selecting the appropriate trees number of feature, and number of training
for further processing. The ensuing image, training time, accuracy and memory
problems are how the types of features are requirements. In addition, the reported
constructed, and how the most important performance also depends on the definition
ones from a large feature space are selected. of a “correct” detection result1, 5. Figure 2
More generalized Haar-like features are shows the effects of detection results versus
defined in11, 12
in which the rectangular different criteria, and more discussions can
image regions are not necessarily adjacent, be found in1, 5. The most commonly adopted
and furthermore the number of such method is to plot the ROC curve using the
rectangular blocks is randomly varied11. de facto standard MIT + CMU data set 1

Several greedy algorithms have been which contains frontal face images. Another
proposed to select features efficiently by data set from CMU contains images with
exploiting the statistics of features before faces that vary in pose from frontal to side
training boosted cascade classifiers17, 21
. view4. It has been noticed that although the
There are also other fast face detection face detection methods nowadays have
methods that demonstrate promising results, impressive real-time performance, there is
including the component-based face still much room for improvement in terms
4
detector using Naive Bays classifiers , the of accuracy. The detected faces returned by
face detectors using support vector state-of-the-art algorithms are often a few
7, 22, 23 24
machines , the Anti-face method pixels (around 5) off the “accurate”

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 65

locations, which is significant as face images alone. The research will focus on
images are usually standardized to 21 by 21 improvement of detection precision for face
pixels. While such results are the trade-offs exposure.
between speed, robustness and accuracy,
they inevitably degrade the performance of Adaptive Boosting
any biometric applications using the The Adaptive Boosting) is a machine
contents of detected faces. Several post- learning algorithm formulated by Freund
processing algorithms have been proposed and Shapiro14 that learns a strong classifier
to better locate faces and extract facial by combining an ensemble of weak
features (when the image resolution of the classifiers with weights. The discrete
detected faces is sufficiently high)26, 27. Adaptive Boosting algorithm was originally
developed for classification using the
Applications exponential loss function and is an instance
As face detection is the first step of any face within the boosting family.
processing system, it finds numerous
applications in face recognition, face Hear-like features
tracking, facial expression recognition, Similar to the what Haar wavelets are
facial feature extraction, gender developed for basis functions to encode
classification, clustering, attentive user signals, the objective of two-dimensional
interfaces, digital cosmetics, biometric Haar features is to collect local oriented
systems, to name a few. In addition, most of intensity difference at different scale for
the face detection algorithms can be representing image patters. This
extended to recognize other objects such as representation transforms an image from
cars, humans, pedestrians, and signs, etc5. pixel space to the space of wavelet
coefficients with an over-complete
Summary dictionary of features. The Haar-like
The advance in face exposure has created a features, similar to Haar wavelets, compute
lot of exciting and reasonably applications. local oriented intensity difference using
As most of the algorithms can also be rectangular blocks (rather than pixels)
applied to other problem domains, it has which can be computed efficiently with the
broader impact than detecting faces in integral image2.

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 66

of examining all scanned image patches can

Receiver operating characteristic be reduced significantly.
A receiver operating characteristic is a plot
p
commonly used in machine learning and
data mining for exhibiting the performance
of a classifier under different criteria. The y-
y
(a) Face images (b)Non-face images
axis is the true positive and the xx-axis is the
Fig. 1. Four types of Haar
Haar-like features.
false positive (i.e., false alarm). A point on
These features appear at different position
operating characteristic shows
sho that the
and scale.The Haar
Haar-like features are
trade-off
off between the achieved trues
computed as the difference of dark and light
positive detection rate and the accepted
regions. They can be considered as features
false positive rate.
that collect local edge information at
different orientation and scale. The set of
Classifier cascade
Haar-like
like features is large, and only a small
In face detection, a classifier cascade is a
amount of them are learned from positive
degenerate decision tree where each node
and negative
ive examples for face detection.
(decision stump) consists of a binary
classifier. In 2, each node is a boosted
classifier consisting of several weak
classifiers. These boosted classifiers are
constructed so that the ones near the root
(a) Test image (b)Detection results
can be computed very efficiently at very
Fig. 2. Detection results depend heavily on
high detection rate with acceptable false
the adopted criteria. Suppose all the sub-
sub
positive rate. Typically, most patches in a
images in (b) are returned as face patterns
test image can be classified as faces/non
faces/non-
by a detector. A loose criterion may declare
de
faces using simple classifiers near the root,
all the faces as “successful” detections
and relatively few difficult ones need to be
while a more strict one would declare most
analyzed by nodes with deeper depth. With
of them as non
non- faces.
this cascade structure, the total computation

REFERENCES

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 67

1. Rowley, H., Baluja, S., Kanade, T.: 7. Heisele, B., Serre, T., Poggio, T.: A
Neural network-based face detection. IEEE component-based framework for face
Transactions on Pattern Analysis and detection and identification. International
Machine Intelligence 20(1) (1998) 23–38 Journal ofComputer Vision 74(2) (2007)
167–181
2. Viola, P., Jones, M.: Robust real-time
face detection. International Journal of 8. Amit, Y., Geman, D.: A computational
Computer Vision 57(2) (2004) 137–154 model for visual selection. Neural
Computation 11(7) (1999) 1691–1715
3. Sung, K.K., Poggio, T.: Example-based
learning for view-based human face 9. Fleuret, F., Geman, D.: Coarse-to-fine
detection. IEEE Transactions on Pattern face detection. International Journal of
Analysis and Machine Intelligence 20(1) Computer Vision 41(12) (2001) 85–107
(1998) 39–51
10. Papageorgiou, C., Poggio, T.: A
4. Schneiderman, H., Kanade, T.: Object trainable system for object recognition.
detection using the statistics of parts. International Journal of Computer Vision
International Journal of Computer Vision 38(1) (2000) 15–33
56(3) (2004)151–177
11. Dollar, P., Tu, Z., Tao, H., Belongie, S.:
5. Yang, M.H., Kriegman, D., Ahuja, N.: Feature mining for image classification. In:
Detecting faces in images: A survey. IEEE Proceedings of IEEE Conference on
Transactions on Pattern Analysis and ComputerVision and Pattern Recognition.
MachineIntelligence 24(1) (2002) 34–58 (2007)

6. Mohan, A., Papageorgiou, C., Poggio, T.: 12. Li, S., Zhang, Z.: Floatboost learning
Example-based object detection in images and statistical face detection. IEEE
by components. IEEE Transactions on Transactions on Pattern Analysis and
PatternAnalysis and Machine Intelligence Machine Intelligence28(9) (2004) 1112–
23(4) (2001) 349–361 1123

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 68

13. Huang, C., Ai, H., Li, Y., Lao, S.: High- object detection. In: Proceedings of IEEE
performance rotation invariant multiview Conference on Computer Vision and Pattern
face detection. IEEE Transactions on Recognition. (2007)
PatternAnalysis and Machine Intelligence
29(4) (2007) 671–686 19. Brubaker, S.C., Wu, J., Sun, J., Mullin,
M., Rehg, J.: On the design of cascades of
14. Freund, Y., Schapire, R.: A decision- boosted ensembles for face detection.
theoretic generalization of on-line learning International Journal of Computer Vision
and application to boosting. Journal of 77(1-3) (2008) 65–86
computer andsystem sciences 55(1) (1997)
119–139 20. Provost, F., Fawcett, T.: Robust
classification for imprecise environments.
15. Jones, M., Viola, P.: Fast multi-view Machine Learning 42(3) (2001) 203–231
face detection. Technical Report TR2003-
96, Mitsubishi Electrical Research 21. Pham, M.T., Cham, T.J.: Fast training
Laboratories (2003) and selection and Haar features using
statistics in boosting-based face detection.
16. Viola, P., Jones, M.: Fast and robust In: Proceedings of IEEE International
classification using asymmetric Adaboost Conference on Computer Vision. (2007)
and a detector cascade. In: Advances in
Neural Information Processing Systems. 22. Oren, M., Papageorgiou, C., Sinha, P.,
(2002) 1311–1318 Osuna, E., Poggio, T.: Pedestrian detection
using wavelet templates. In: Proceedings of
17. Wu, J., Brubaker, S.C., Mullin, M., IEEE Conference on Computer Vision and
Rehg, J.: Fast asymmetric learning for Pattern Recognition. (1997) 193–199
cascade face detection. IEEE Transactions
on Pattern Analysis and Machine 23. Romdhani, S., Torr, P., Sch¨olkopf, B.,
Intelligence 30(3) (2008) 369–382 Blake, A.: Computationally efficient face
detection. In: Proceedings of the Eighth
18. Pham, M.T., Cham, T.J.: Online IEEE International Conference on
learning asymmetric boosted classifiers for

http://www.srji.co.cc
 Vol.1 ● No.3 ● 2012 Scientific Research Journal of India 69

Computer Vision. Volume 2. (2001) 695– model and its application to analysis of
700 facial images. IEEE Transactions on Pattern
Analysis and Machine Intelligence 5(28)
24. Keren, D., Osadchy, M., Gotsman, C.: (2006) 73800752
Antifaces: A novel fast method for image
detection. IEEE Transactions on Pattern 27. Ding, L., Martinez, A.: Precise detailed
Analysis and Machine Intelligence 23(7) detection of faces and facial features. In:
(2001) 747–761 Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition.
25. Osadchy, M., LeCun, Y., Miller, M.: (2008)
Synergistic face detection and pose
estimation with energy-based models. 28. Friedman, J., Hastie, T., Tibshirani, R.:
Journal of Machine Learning Research Additive logistic regression: a statistical
(2007) 1197–1214 view of boosting (With discussion and a
rejoinder by the authors). The Annals of
26. Moriyama, T., Kanade, T., Xiao, J., Statistics 28(2) (2000) 337–407
Cohn, J.: Meticulously detailed eye region

CORRESPONDENCE
*Centre for Research and Development. PRIST University, India. E-Mail:svthanigaivel@gmail.com

http://www.srji.co.cc