ROBUST FACE DETECTION IN PATIENT TRIAGE IMAGES

iyati Chhaya and Tim Oates

University of Maryland, Baltimore County, 1000 Hilltop Circle, Baltimore, Maryland, U.S.A.

Keywords:

Face detection, Feature extraction, Triage images.

Abstract:

Disaster events like the attack on the World Trade Center in New York City in 2001 and the earthquake in Haiti

in 2010 result in a desperate need for family and friends to obtain information about missing people. This

can be facilitated by automatically extracting textual descriptors of patients from images taken at emergency

triage centers. We show that existing face detection algorithms, a necessary precursor to extracting such

descriptors, perform poorly on triage images taken during disaster simulations, and propose and demonstrate

the superiority of an ensemble-based face detection method that is a combination of robust skin detection and

simple pattern matching face detection techniques. The addition of a template-based eye detection method

further improves the accuracy of our face detection method.

1 INTRODUCTION

Disaster events like the attack on the World Trade

Center in New York City in 2001 and the earthquake

in Haiti in 2010 have both local and non-local effects.

Locally, there is signiﬁcant loss of life and property.

Non-locally, there is a desperate need for informa-

tion about the status of friends and family who might

have been directly impacted. An increasingly com-

mon practice in disaster preparedness drills, which are

required annually of hospitals in the United States,

is the use of digital cameras to take pictures of pa-

tients as they are triaged. Privacy concerns, as well

as the sometimes graphic nature of the images when

injuries are severe, makes it impractical to make the

images publicly accessible. However, descriptive fea-

tures extracted from the images – hair color and style,

clothing color and style, distinctive features like tat-

toos, gender, ethnicity, age, etc. – can be made pub-

licly available and searchable.

To extract descriptive features from triage images,

one must ﬁrst locate the patient’s face. This proves to

be a difﬁcult task given that triage images are taken

in highly variable lighting conditions from many dif-

ferent viewing angles and distances, and the patients’

faces are often obstructed with bandages, dirt, debris,

and blood. In this paper we propose a novel ensem-

ble based face detection algorithm that combines skin

detection and face detection based on pattern match-

ing that works well (in terms of precision and recall)

for our corpus of images. We also show that state-of-

the-art face detection algorithms perform poorly on

a corpus of patient triage images taken during a dis-

aster preparedness drill. Finally, we apply template

based eye detection along with the proposed face de-

tection algorithm to show how locating eyes in the

image gives better results for the resultant face detec-

tion. Figure 1 shows the overview of the proposed

ensemble based algorithm.

The remainder of the paper is organized as fol-

lows. Section 2 reviews related work for skin, face,

and eye detection. Our proposed approach is de-

scribed in Section 3. Section 4 presents our dataset

and empirical results, and Section 5 concludes and

points to future work.

2 RELATED WORK

Even in the domain of disaster response where faces

can be obstructed in a variety of ways, good skin de-

tection is essential. Probabilistic models such as (El-

gammal et al., 2009) compute the probability that a

pixel corresponds to skin given, for example, it’s RGB

values. Others approaches (Vassili et al., 2003) make

hard decisions. For example, Peer et al. (Peer et al.,

2003) declare all pixels that meet the following con-

ditions to correspond to skin:

R > 95, G > 40, B > 20

max(R, G, B) − min(R, G, B) > 15

|R− G| > 15, R > G, R > B

(1)

326

Chhaya N. and Oates T..

ROBUST FACE DETECTION IN PATIENT TRIAGE IMAGES.

DOI: 10.5220/0003319303260333

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2011), pages 326-333

ISBN: 978-989-8425-47-8

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

!"#$%#&#'()*$

+,-$./0/1#&#02$

3/4#0*$5/&'67*8$

9/'#$%#&#'()*$

:#1.;/&#$5/&'67*8$

<=7*$%#&#'()*$$

Figure 1: An overview of the processing done on each patient image for face detection.

This algorithm is a component of ours. Once relevant

pixels are found, they can be grouped into distinct re-

gions that may correspond to different body parts or

different people (Aznaveh et al., 2008; Ravichandran

and Ananthi, 2009; Phung et al., 2005).

In this work we are concerned with face region de-

tection (Yang et al., 2002) (ﬁnding contiguous regions

in an image that mostly contain faces), which is con-

trasted with face recognition (determining the identity

of a person based on an image of their face) and face

detection (marking the locations of faces, but not their

spatial extent). Some approaches use combinations of

facial features for face region localization. For exam-

ple, (Pai et al., 2006), which we incorporate in our al-

gorithm, uses the height to width ratio of the face and

the presence of eyes and a mouth. Ratios in the range

[0.75, 1.75] are acceptable, and eye and mouth loca-

tions are found using color histograms and shape fea-

tures. The baseline against which we compare our re-

sults is the Viola-Jones face detection algorithm (Vi-

ola and Jones, 2001) which uses a feature representa-

tion based on Haar wavelets and trains a discrimina-

tive model using AdaBoost (Collins et al., 2000). This

algorithm has state-of-the-art performance on com-

monly used datasets.

Many face detection algorithm can take advan-

tage of information on the location of eyes. There

are numerous approaches to eye localization, such as

the use of active IR (Peng et al., 2005) which is not

feasible for disaster events due to the requirement for

specialized equipment. Others use eye templates ex-

tracted from images either to train supervised clas-

siﬁers or for matching. We incorporate a template

matching algorithm in our approach. Combination

approaches such as locally selective projection with

SVMs (Zheng and Wang, 2008), rotation invariant

Zernike moments with SVMs (Kim and Kim, 2008),

and face circle ﬁtting with dark pixel ﬁlters (Lin and

Yang, 2004) have also been studied.

Interestingly, many of the most commonly used

datasets in face detection were originally developed

for face recognition, and thus are relatively uni-

form. For example, the FERET face dataset (Phillips

et al., 2000) contains images taken from frontal and

left/right proﬁle views, the Yale dataset (Yal, 2006)

contains frontal images with glasses and various fa-

cial expressions, and the MIT dataset (MIT, 2000)

contains images of 10 different people taken under

different lighting conditions and poses. All of these

datasets lack diversity along a number of dimensions:

race, open vs. closed eyes, presence of facial hair, tat-

toos, etc. This is part of the reason that existing al-

gorithms, developed and tested on uniform data, per-

form poorly on images from disaster events that vary

along all of these dimensions and many others.

ROBUST FACE DETECTION IN PATIENT TRIAGE IMAGES

327

3 PROPOSED FACE DETECTION

ALGORITHM

As shown in Figure 1, the input patient image un-

dergoes a sequence of processing events. The face

detection task is treated as a combined result from

a color–based skin detector, pattern matching–based

face detector and a template matching–based eye de-

tector. The system takes an input image and cleans up

the background using skin detection, this processed

image is then fed into the face detection algorithm.

Using this combination the system needs to ﬁnd the

face only within the detected skin region. To fur-

ther improve face detection accuracy, eye detection

is used. After the skin detection step, eye detection

is performed, this gives us eye coordinates. This out-

put is fed into the face detection algorithm and then

pattern matching is done with the already known eye

coordinates. The system is hence a combination of

interlocked algorithms where the feedback from eye

detection is used to improve the performance of the

face detection algorithm. Each component of this en-

semble system is explained here.

Referring back to Figure 1, ﬁrst, skin detection

is done. The skin detection algorithm of Peer et

al. (Peer et al., 2003) is used, modiﬁed slightly

so that max(R, G, B) − min(R,G,B) > 10 rather than

max(R, G, B) − min(R, G, B) > 15 to allow for more

variation in ethnicity, obstruction by dirt and blood,

etc.

As mentioned, skin detection allows us to elimi-

nate the background and focus on those regions that

are most likely to contain a face, which is done with

the face detection algorithm in (Pai et al., 2006) that

uses the height to width ratio of the skin region and

the presence and location of eyes and a mouth. If any

of these features are present in the image the algo-

rithm computes a bounding rectangle around the face.

We relaxed the maximum allowed height to width ra-

tio from 1.75 to 1.8 to allow for more variation. The

eye localization algorithm uses color histograms in

the Y

′

color space. Likewise for mouth local-

ization.

For eye detection, we implemented a template-

based eye detection algorithm given that many of

the patients have their eyes closed, thus confound-

ing color-based approaches. Fifty templates were ex-

tracted from images not used in our evaluation. Some

of them are shown in Figure 2. For matching we

use two different methods and combine the results.

The ﬁrst is the normalized sum of squared differ-

ences (SSD) between the template and image patches;

the second is the normalized cross correlation (NCC).

NCC is useful when lighting conditions vary, as they

do in our images. Equation 2 and Equation 3 give

expressions for SSD and NCC, respectively.

ssd(u, v) = Σ

x,y

( f (x, y) −t(x− u, y− v))

(2)

ncc(u, v) =

n− 1

x,y

( f (x, y) −

f)(t(x, y) −

(3)

In the above equations, n is the number of pixels,

f is the image, t is the template, and the summation is

over positions x, y under the template at position u, v.

σ is the standard deviation.

(a) (b) (c)

Figure 2: Example eye templates.

Because patients can be, and are, in many differ-

ent orientations in our images, we rotate each eye

template and use the rotated templates as indepen-

dent templates for comparison. The rotation is done

around the center of the eye template in 30 degree in-

crements, producing twelve different templates from

each original. Other transformations, such as scal-

ing, were tested, but had little impact on the results.

Empirically, the template-based method works better

than the built-in method of (Pai et al., 2006), so to im-

prove face localization we ﬁrst locate the eyes using

templates and then run the face detection algorithm.

This system being an ensemble of algorithms,

treats the face as a combination of different features

on a given skin area, whereas the Viola-Jones ap-

proach, the system is looking for a complete set of

Haar-like features that match the face. Hence, if there

is a partial face or a face with a different orientation,

our system performs considerably better than the stan-

dard face detection algorithms that are trained spe-

cially for full upright straight face images.

4 EXPERIMENTAL RESULTS

Recall that our problem domain is images of pa-

tients taken during disaster events. Much of our data

came from drills conducted by the Bethesda Hospi-

tals’ Emergency Preparedness Partnership (BHEPP),

which is a collaboration between the National Navel

Medical Center, the National Institutes of Health

Clinical Center, and the Suburban Hospital in

Bethesda, Maryland. One of the goals of BHEPP is

to develop a transportable, proven model for deﬁning

VISAPP 2011 - International Conference on Computer Vision Theory and Applications

328

Table 1: Dataset Summary.

Source Count Remark

CMAX 22 BHEPP disaster drill

Web 6

Lost Person Finder 22 Female

Lost Person Finder 26 Male

Lost Peron Finder 21 Child

Total 97

best practices in casualty care management through

ongoing multi-agency collaboration, integration, and

training. Part of the training is an annual Collabora-

tive Multi-Agency Exercises (CMAX) drill in which a

disaster event is simulated, complete with actors play-

ing the role of disaster victims. Patients are taken to

emergency medical care centers for treatment where

they are triaged and a primary record is created, in-

cluding an image of the patient. In addition, some

of our images come from the National Institutes of

Health Lost Person Finder (LPF) project, which is a

web-based system for re-uniting family and friends

with victims during disaster events.

Figure 3 shows sample images from our dataset

and contrasts them with images often used in face de-

tection and recognition work. Note that the faces in

the BHEPP images are in different orientations, ﬁll

different amounts of the ﬁeld of view, are obstructed

by bandages and other visible signs of injury, have

open and closed eyes, are surrounded by other pa-

tients and medical personnel, etc. In contrast, images

from standard face datasets are remarkably uniform.

As we will see shortly, state-of-the art face detec-

tion algorithms perform exceptionally poorly on the

BHEPP images, necessitating a new approach. Table

1 gives a summary of the 97 images used in our ex-

periments.

We compared four different face detection ap-

proaches as described below:

• OpenCV: The state-of-the-art Viola-Jones face

detection algorithm (Viola and Jones, 2001) im-

plemented in the OpenCV toolkit

• Skin + OpenCV: Skin detection is done before

running the Viola-Jones algorithm

• Skin + FaceDetect: Skin detection is done before

running the face detection algorithm of Pai et al.

(Pai et al., 2006)

• Skin + Eye + FaceDetect: This is the same as Skin

+ FaceDetect except the template-based eye de-

tection method is used

For setting a standard testing platform, we asked

humans to manually annotate our image set for face

detection. They were asked to draw the tightest

bounding rectangle around the face region in each im-

age. Figure 4 shows examples of these annotated im-

ages. These annotations were used as reference im-

ages and results from each of the different methods

are compared against the annotated reference images.

For numerical analysis, we use precision/recall values

and the F-Score, deﬁned as:

Precision(P) =

Annotated ∩ Program

Program

Recall(R) =

Annotated ∩ Program

Annotated

F − Score =

2 × P × R

P + R

(4)

We calculate precision and recall values by count-

ing the amount of pixel overlap between the human

annotations and the bounding boxes produced by the

algorithms. If the result from our algorithm falls com-

pletely inside the human-annotated area then we have

perfect precision and a lower recall value. If the result

from our algorithm completely contains the human-

annotated rectangle then we have perfect recall and

lower precision. And, ﬁnally, if there is no overlap

between the regions from human annotation and that

from program then precision and recall are 0. There

is no distinction between no overlap and no detection

(no rectangle marked) in the results.

4.1 OpenCV

On testing face detection using OpenCV on our im-

ages we got extreme results in most cases. Large

numbers of images had no detection of the face (pre-

cision/recall=0) whereas some images had perfect de-

tection of the face (precision/recall=1). Table 2 sum-

marizes the results. Due to OpenCV’s failure to de-

tect faces in almost 50% of the images, the average

F-Score is also low (0.47) for this algorithm.

Table 2: Results: OpenCV.

Property Number of Images

Total Images 97

No Detection 48

Perfect Detection 28

Average Recall 0.49

Average Precision 0.47

Average F-Score 0.47

The images with perfect face detection are all

frontal clean face images, very similar to those from

ROBUST FACE DETECTION IN PATIENT TRIAGE IMAGES

329

(a) Source: BHEPP (b) Source: BHEPP (c) Source: BHEPP

(d) OpenCV test image:

Lena

(e) Images from Yale face database (f) Proﬁle test image from

CMU face database

Figure 3: Images from BHEPP disaster drills (a - c) compared to images often used in face detection and recognition work (d

- f).

(a) (b) (c)

Figure 4: Annotated Images.

Table 3: Results: Skin + OpenCV.

Property Number of Images

Total Images 97

No Detection 37

Perfect Detection 31

Average Recall 0.6

Average Precision 0.58

Average F-Score 0.59

standard face databases. As OpenCV is trained us-

ing Haar wavelets for frontal face features it fails for

anything outside that frame. Cascading of Haar fea-

tures, done as a part of the OpenCV system, also

does not improve performance for non-frontal im-

ages. Another observation is that a cluttered back-

ground greatly affects the performance of OpenCV

face detection. In some images, non-face regions

were detected as faces.

4.2 Skin + OpenCV

The performance of OpenCV showed us that chang-

ing the background distracts the algorithm. Our next

experiment involved ﬁrst doing skin detection and

then continuing to run OpenCV face detection on the

images with only the skin area visible. This method

helped us eliminate background noise. As seen from

the results, the overall performance of the algorithm

improved. Also the number of no-detections went

down from 48 images in the previous experiment to

37 images in this experiment. Table 3 summarizes

these results.

4.3 FaceDetect

We ﬁrst performed skin detection on all images and

then applied the pattern–matching–based face detec-

tion algorithm (FaceDetect) on the detected areas.

The results in Table 4 show that there were very

VISAPP 2011 - International Conference on Computer Vision Theory and Applications

330

Table 4: Results: Skin + FaceDetect.

Property Number of Images

Total Images 97

No Detection 7

Perfect Detection 0

Average Recall 0.66

Average Precision 0.72

Average F-Score 0.65

Table 5: Results: Skin + Eye + FaceDetect.

Property Number of Images

Total Images 97

No Detection 1

Perfect Detection 1

Average Recall 0.77

Average Precision 0.83

Average F-Score 0.78

few images with no face detection at all. Though

at the same time, the number of images with perfect

face detection is zero. The more frontal face images

give high recall values but the precision is not per-

fect. Also, we observe that most face detected re-

gions cover only part of the actual face and tend to

be towards one side of the face rather than covering

the complete face area.

4.4 Eye + FaceDetect

The results of face detection from the previous sec-

tion showed that the detected face tends to be shifted

on either side of the actual face. In this experiment,

we ﬁrst implemented eye detection on detected skin

areas and then did the face detection with the eye lo-

cations known. Table 5 summarizes the results for

all images. The number of no–detections reduces to

just one. Also, the average precision/recall value is

very high as compared to that from prior results. The

increase in average F-Score shows that there is an im-

provement in precision, recall, or both.

4.5 Comparisons

In this section we compare the results of all the above

methods. Figure 5 shows results when only OpenCV

is used and when OpenCV is used after skin detec-

tion. From the graph we observe that the average per-

formance improves and the number of no-detections

also decreases.

Figure 6 shows results from both methods using

the pattern–matching–based face detection. As seen

from the graph, average performance for both of these

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

OpenCV, Skin + OpenCV

Recall

Precision

OpenCV

Average: OpenCV

Skin + OpenCV

Average: Skin + OpenCV

Figure 5: Graph: OpenCV, Skin + OpenCV – Each point

in the graph represents precision/recall value of one image

from the database.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Skin + Facedetect, Skin + Eye + Facedetect

Recall

Precision

Skin + Facedetect

Average: Skin + Facedetect

Skin + Eye + Facedetect

Average: Skin + Eye + Facedetect

Figure 6: Graph: Skin + FaceDetect, Skin + Eye + FaceDe-

tect – Each point in the graph represents precision/recall

value of one image from the database.

methods is better than OpenCV and also that the av-

erage after Eye + FaceDetect improves as compared

to that of FaceDetect.

Table 6: Results: Average F-Score Value.

Method Avg. F-Score value

OpenCV 0.47

Skin + OpenCV 0.59

Skin + Facedetect 0.65

Skin + Eye + Facedetect 0.78

4.6 Eye Detection

Figure 8 shows the detected face area for a given im-

age using different methods. The complete rectan-

gle is the annotated reference region, the – rectangle

is from FaceDetect result and the -. rectangle shows

ROBUST FACE DETECTION IN PATIENT TRIAGE IMAGES

331

Table 7: Results: Eye Detection–Without template rotation.

Image Count Correct Detected(%) False Positives No detected Templates

Open Eyes 67 51 (76.12) 13 3 8

Closed Eyes 10 4 (40.00) 1 5 3

Half Open 15 7 (46.67) 1 7 4

Not Visible 5 0 (-) 1 0 0

All 97 62 (63.91) 16 15 15

Table 8: Results: Eye Detection–With template rotation.

Image Count Correct Detected (%) False Positives Not detected Templates

Open Eyes 67 58 (86.57) 8 1 6

Closed Eyes 10 8 (80.00) 1 1 3

Half Open 15 14 (93.33) 1 0 3

Not Visible 5 0 (-) 1 0 0

All 97 80 (82.48) 11 2 12

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Average: All methods

Recall

Precision

OpenCV

Skin + OpenCV

Skin + Facedetect

Skin + Eye + Facedetect

Figure 7: Graph: All Methods - Average.

Figure 8: Face Detection Result: — : Human annotated - -

- : Skin + FaceDetect .-: Skin + Eye + FaceDetect.

result of the Eye + FaceDetect algorithm. As seen

from the image, FaceDetect detects partial face and

FaceDetect done after eye detection detects a more

accurate face region. These results are reﬂected in

Figure 7 where average precision and recall from all

methods are shown. The graph shows an almost lin-

ear improvement in performance. Table 6 compares

average F-Scores from all methods. These numbers

also show improvement as the methods change. The

best performance is shown using the Skin + Eye +

FaceDetect method.

We performed an accuracy test to test the eye de-

tection algorithm by itself. The images under the

non–visible eyes section are those images where the

person is either wearing sunglasses or has a bandage

around the eye region.

Tables 7 and 8 show results of eye detection using

template matching with normal templates and using

rotated templates respectively. As observed from the

results, eye detection with template rotation performs

better than that without any rotation. The principal

reason behind this result is the presence of large num-

ber non–upright face images in the dataset.

5 CONCLUSIONS

We began with comparing standard face databases

against the dataset from real life mass disaster situ-

ation. The results from the previous section show that

an ensemble–based face detection algorithm performs

better than using off–the–shelf standard face detec-

tion algorithms. Relaxing parameters for skin detec-

tion helps in accommodating a wider range of skin

color and noise in the skin area, hence resulting in

improved performance. Performance of face detec-

tion algorithms improves drastically with elimination

of unwanted background regions. Also, the face de-

tection algorithm that we use gives better accuracy if

eye detection is done ﬁrst. Template–matching–based

eye detection using a normalized sum of squared dif-

VISAPP 2011 - International Conference on Computer Vision Theory and Applications

332

ference and normalized cross correlation is effective

for this database. The key aspect of the algorithm is

rotation of templates for comparison. The rotation al-

lows accurate detection of eyes in images where the

person is lying down or not in an upright position.

Hence, we have both face to eye detection and eye

to face detection. This allows us to check the perfor-

mance of each algorithm and use the other one as a

feedback mechanism to improve performance.

We developed a mesh–like structure with the al-

gorithms, where each algorithm acts as a feedback for

the other, hence validating the performance and aid-

ing to improve performance of individual algorithms.

Our work also illustrates how algorithms meant for

standard face databases can be modiﬁed for real–life

noisy databases. We use individual unrelated methods

at each stage of detection such as color model analy-

sis for skin detection, pattern–matching for face de-

tection and template–matching for eye detection, all

of these ﬁnally coming together for an improved face

detection algorithm.

REFERENCES

2000, MIT Face Recognition Database. Online.

2006, Yale Face Database. Online.

Aznaveh, M. M., Mirzaei, H., Roshan, E., and Saraee, M.

(2008). A new color based method for skin detection

using rgb vector space.

Collins, M., Schapire, R. E., and Singer, Y. (2000). Lo-

gistic regression, adaboost and bregman distances. In

Conference on Learning Theory, pages 158–169.

Elgammal, A. M., Muang, C., and Hu, D. (2009). Skin de-

tection. In Encyclopedia of Biometrics, pages 1218–

1224. Springer US.

Kim, H.-J. and Kim, W.-Y. (April 2008). Eye detection

in facial images using zernike moments with svm.

Electronics and Telecommunications Research Insti-

tute, 30:335–338.

Lin, D.-T. and Yang, C.-M. (2004). Real-time eye detec-

tion using face-circle ﬁtting and dark-pixel ﬁltering.

In International Conference on Multimedia and Expo,

pages 1167–1170.

Pai, Y.-T., Ruan, S.-J., Shie, M.-C., and Liu, Y.-C. (2006). A

simple and accurate color face detection algorithm in

complex background. IEEE International Conference

on Multimedia and Expo, 0:1545–1548.

Peer, P., Kovac, J., and Solina, F. (2003). Human skin color

clustering for face detection.

Peng, K., Chen, L., Ruan, S., and Kukharev, G. (2005).

A robust and efﬁcient algorithm for eye detection on

gray intensity face. In International Conference on

Advances in Pattern Recognition (2), pages 302–308.

Phillips, P. J., Moon, H., Rizvi, S. A., and Rauss, P. J.

(2000). The feret evaluation methodology for face-

recognition algorithms. IEEE Transactions on Pat-

tern Analysis and Machine Intelligence, 22(10):1090–

1104.

Phung, S. L., Bouzerdoum, A., and Chai, D. (2005). Skin

segmentation using color pixel classiﬁcation: Anal-

ysis and comparison. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 27(1):148–154.

Ravichandran, K. and Ananthi, B. (2009). Color skin

segmentation using k-means cluster. International

Journal of Computational and Applied Mathematics,

4:153–157.

Vassili, V. V., Sazonov, V., and Andreeva, A. (2003). A

survey on pixel-based skin color detection techniques.

In in Proc. Graphicon-2003, pages 85–92.

Viola, P. and Jones, M. (2001). Robust real-time object de-

tection. In International Journal of Computer Vision.

Yang, M.-H., Kriegman, D. J., and Ahuja, N. (2002). De-

tecting faces in images: A survey. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

24(1):34–58.

Zheng, Y. and Wang, Z. (2008). Robust and precise eye

detection based on locally selective projection. In In-

ternational Conference on Pattern Recognition, pages

1–4.

ROBUST FACE DETECTION IN PATIENT TRIAGE IMAGES

333