-
Eyes Detection in Facial Images using Circular Hough
Transform
W. M. K Wan Mohd Khairosfaizal and A. J. NorainiFaculty of
Electrical Engineering,
Universiti Teknologi Mara,40450 Shah Alam, Selangor,
Malaysia
Abstract-This paper presents an eye detection approach using
Circular Hough transform. Assuming the face region has already been
detected by any of the accurate existing face detection methods,
the search of eye pair relies primarily on the circular shape of
the eye in two-dimensional image. The eyes detection process
includes preprocessing that filtered and cropped the face images
and Circular Hough Transform is used to detect the circular shape
of the eye and to mark the eye pair on the image precisely. This
eyes detection method was tested on Face DB database developed by
Park Lab, University of Illinois at Urbana Champaign USA. Most of
the faces are frontal with open eyes and some are tilted upwards or
downwards. The detection accuracy of the proposed method is about
86%.
Keyword- Accumulation array, Gradient magnitude, Gradient
thresholding, Circular Hough Transform
I. INTRODUCTION
Human eyes play an important role in face recognition and facial
expression analysis. In fact, the eyes can be considered salient
and relatively stable feature on the face in comparison with other
facial features. Eye detection is valuable in determining the
orientation of the face and also the gaze direction. The position
of other facial features can be estimated using the eye position
[1]. In addition, the size, the location and the image-plane
rotation of face in the image can be normalized by only the
position of both eyes. This is also regarded as one of the most
important biometrics characteristics for personal
identification.
The existing work in eye position detection can be classified
into two categories. First, the active infrared (IR) based
approaches and second the image-based passive approaches. Eye
detection based on active remote IR illumination is a simple yet
effective approach [2]. But it relieson an active IR light source
to produce the dark or bright pupil effects. In other words, this
method can only be applied to the IR illuminated eye images. This
method is not widely used, because in many real applications the
face images are not IR illuminated.
The image-based passive methods can be classified into three
categories. First, template based method [3-6], secondly is the
appearance based method [7-9] and the third is feature based method
[10-14]. In the template based method, a generic eye model, based
on the eye shape, is designed first. Template matching is then used
to search the image for the eyes. While this method can detect eyes
accurately, it is normally time-
consuming. The appearance based method detects eyes based on
their photometric appearance. This method usually needs to collect
a large amount of training data, representing the eyes of different
subjects, under different face orientations, and under different
illumination conditions. These data are used to train a classifier
such as a neural network or the support vector machine and
detection is achieved via classification. Feature based methods
explore the characteristics such as edge and intensity of iris, the
color distributions of the sclera and the flesh of the eyes to
identify some distinctive features around the eyes. Although this
method is usually efficient, they lack of accuracy for the images
which do not have high contrast. For example, these techniques may
have mistaken eyebrows for eyes.
In this paper, the face detection is carried out on the
identified face region without detecting the face region first as
the main focus is to detect eye pair from the face image. A simple
yet robust algorithm to locate the eye pair on grey intensity face
images is proposed. Currently, there are lots of promising face
detection methods [15-18] exist thus assumptions have been made in
this work such that (1) a rough face region has been located, (2)
the image consists of only one face and (3) eyes in face image can
be seen. The image-based eye detection approaches is used to locate
the eyes by exploiting eyes differences in appearance and shape
from the rest of the face. The special characteristics of the eye
such as dark pupil, white sclera, circular iris, eye corners, eye
shapeand etcetera are utilized to distinguish the human eyes from
other objects. The steps involve in the eye detection process are
cropping the face images to the required face region, threshold on
the gradient magnitude of the face images to get the linear indices
in the images and since the iris is nearly circular, the Hough
transform is used to detect the circular shape of the iris of the
human eye based on the linear indices. The pupil of the eye is
plotted as the circle center and the circular shape of the iris is
located and drawn as the circle parameter with its specific radius
from the circle center. The proposed method is expected to increase
the efficiency of feature based methods.
II. METHODOLOGY
The block diagram of the proposed approach for the eye detection
is shown in Figure 1.
238
2009 5th International Colloquium on Signal Processing & Its
Applications (CSPA)
978-1-4244-4152-5/09/$25.00 2009 IEEE
-
The process of detecting the eye pair in the face image starts
with acquiring the grey scale face image from the face database.
The image must be two dimensions with the rough face region
consists of a face and eyes. The algorithm built can only be used
under this situation. The output image is known as the raw image.
Face detection will process first locate the rough face region. In
the second stage, an efficient feature-based method is used to
locate two rough regions of the eyes in the face, which is the
objective of the study.
A. Preprocessing In order to obtain a proper segmentation of the
image, pre-processing of the image is carried out. To compensate
for illumination variations and to obtain more image details, a
median filter is used to enhance the brightness and the contrast of
the images [20]. It is also used to eliminate the noise from the
raw image. A median filter is based upon moving a window over an
image and computing the output pixel as the median value of the
brightness within the input image. A useful variation on the theme
of the median filter is the percentile filter. Here the centre
pixel in the window is replaced not by the 50% (median) brightness
value but rather by the p% brightness value where p% ranges from 0%
(the minimum filter) to 100% (the maximum filter). Values other
then (p=50) % do not, in general, correspond to smoothing filters.
This step simultaneously normalizes the brightness across an image
and increases contrast. As a result, the image is enhanced and
corrected from noise. The face region from the filtered image is
cropped out from the background. This is done to eliminate the
unwanted region and also to facilitate the process of detecting the
eyes. The output image from this stage is known as the filtered
image.
B. Eye Pair Detection When the rough face region is detected,
the eye pair detection is sequentially applied to locate the rough
regions of both eyes. Figure 2 shows the process of the proposed
method.
C. Validation of Image Parameter. This step is to validate the
filtered image parameters in order to ensure that the subsequent
algorithms used can be applied. The parameters that need to be
considered are as follow,
i. Dimension (2-D)ii. Size (minimum 32X32)
iii. Type ( greyscale image)
D. Building the Accumulation Array To build the accumulation
array, the first step is to compute the gradient and the gradient
magnitude of the roughface image region. It is the first derivative
of two-dimensional image. The equations used are as follow:
i. Two dimensional first derivative; (1)
where; xh :denotes a horizontal derivative, yh :denotes a
vertical derivative,
h :denotes the arbitrary angle derivative.
ii. Gradient, a[m,n], of an image:
yyxxy iahiahi
y
aix
x
aa
(2)where ix and iy are unit vectors in the horizontal and
vertical direction, respectively.
iii. Gradient magnitude,
2y2x ahaha (3)
FACE IMAGE
FILTER(Median Filter)
CROP(Extacting the Face
Region)
Eye Pair Detection
Preprocessing
Fig. 1. Block Diagram of the Eye Detection Process
yx hsinhcosh
Validation of Parameter(Accumulator Building)
Area of Interest
Hough Transform to Detect Eye
(Circular Hough Transform)
Fig. 2. Block Diagram of the Eye Pair Detection Process
239
-
Approximated by;
ahaha yx ~ (4)
The linear indices of the gradient magnitude are computedusing
the equation as follows;
jijn1j
kik Xaxf (5)
where; ija :gradient magnitude,
jX : Symmetry square matrix,
ik xf : Linear indices of the gradient magnitude.
The accumulation array of the image consists of the gradient
magnitude of the image and its linear indices as in equation
(6).
Accumulator = (Gradient Magnitude, Linear Indices) (6)
E. Area of Interest From the segmentation process of the
accumulator, the segmented accumulator is smoothened to get better
segmented value using averaging filter. To obtain the area of
interest, local maxima mapping on the image face region is
generated by locating every local maximum on the segmented region.
Local maximum filter is built by thresholding the local maxima
mapping with the lower bound value. There are twosteps to be done
separately. First, the segmented accumulator is threshold with the
non-segmented accumulator and filtered by the local maximum filter.
Next step is to label the generated local maximum mapping by eight
connected component as in Figure 3 and then threshold by the
gradient component. The threshold process is known here by gradient
threshold and basically takes the adaptive threshold method which
threshold value varies across the entire image. The equation of the
threshold method is as in Equation (7). The output from the second
step is known as mask.
Fig. 3. Label 8 Connected Component
cffTT , (7) where; T is the threshold
f is the whole image,
cf is 8 label image part.
The results from both steps are compared to select the area of
interest in the face image. The comparison of both value and the
reconstructed data image are as in Figures 4 and 5respectively:
Fig. 4. Comparison Graph of values from steps 1 and 2
Fig. 5. Reconstructed Image
The reconstructed image indicates which area in the image can be
considered as the area of interest (location of eye pair). The
clipping value as seen in the graph is threshold by the gradient
magnitude values at the respective area. The respective values in
the accumulator array are replaced by these new threshold values.
Then the process of locating the local maxima on every threshold
area is done to detect the eye area in the image. As the
accumulator array of the reconstructed image can be converted into
a function of f(x), then the local maximum is at xo for a>0 such
that, for x (xo-a, xo +a) there is f(x)f(xo). Intuitively, it means
that around xo the graph of f will be below f(xo). Each area of
interest is compiled together as further process is done only among
these components. Each local maxima candidate in every area of
interest is compiled in a group by selecting minimum number of
qualified pixel in each group of the interested area.
F. Circular Hough Transform to Detect Circle of the Eyes Hough
transform is a technique which can be used to isolate features of a
particular shape within an image. Because it requires the desired
features to be specified in some parametric form, the classical
Hough transform is most commonly used for the detection of regular
curves such as lines, circles, ellipses, and etcetera [21]. The
main advantage of the Hough transform technique is that it is
tolerant to gaps in feature boundary descriptions and is relatively
unaffected by image noise. To detect the eye which is circular in
shape the so called Circular Hough Transform is used as in
equation(8).
22020 ryyxx (8)
240
-
where, 00 y,x is the coordinate of the circle centre, r is the
radius of the circle.
The detection process starts with the local maxima in the group
of the area of interest is assumed as the centre of the circle. If
the linear indices among the minimum value of qualified pixel
forming the circular shape, then that area is the eye region
detected on the image. Every area of interest is tested with this
process for it occurs as an element of the circle component which
is the eye region identified in the image.
III. RESULT AND DISCUSSION
The test on using the proposed method was conducted on a well
known Face DB database [22]. The Face DB database has been
developed by the University of Illinois at Urbana-Champaign under
their Productive Aging Laboratory. The laboratory is also known as
the Park Lab, which was named after the founder of the laboratory,
Dr Denise C. Park. In this database there are several types of face
images either colouredimages (RGB) or intensity images
(gray-scale).The image format also varies from Windows Bitmap (BMP)
to Joint Photographic Expert Group (JPEG). The images selected are
in two-dimension with size consists of 646 x 480 gray-scale,BMP
format consist of 72 face images with a constraint background. The
images consist of three ethnics which are the Asians,
African-Americans and Caucasians. The categorizingsets a wider
illumination based on their skin colour. Furthermore, each category
has been subdivided into set of age ranging from 19 to 84 years
old. Out of 72 face images, 50 images have well opened eyes while
the rest have eyes partially opened. The developed software is
written in Matlab codes and the results from the eye detection
process are shown in figures 6 to 12.
Fig. 6. Original image
Fig. 7. Filtered and Cropped Face Image Region.
Generated map of local maxima
50 100 150 200
50
100
150
200
250
Fig. 8. Generated Map on Local Maxima
050
100150
200250
0
100
200
300
-40
-20
0
20
40
Accumulation array after local maximum filtering
Fig. 9. 3-D View of Accumulation Array after Local Maximum
Filtering
Accumulation Array from Circular Hough Transform
50 100 150 200
50
100
150
200
250
300
350
Fig. 10. Accumulation Array from Circular Hough Transform.
050
100150
200250
0
100
200
300
400
0
200
400
600
800
3-D View of the Accumulation Array
Fig. 11. 3-D View of the Accumulation Array from Circular
HoughTransform
Grayscale Image with Detected Eye Circle(center positions and
radii marked)
50 100 150 200
50
100
150
200
250
300
350
Fig. 12. The Face Image With the Eyes Pair Detected.
The original face is shown in Figure 6 and Figure 7 shows the
original face that has been cropped to obtain the face region from
the facial image and filtered using median filter. Other than
reduced in noise the filtered image has also been enhanced in term
of its brightness and contrast as to
241
-
compensate for its illumination variations to obtain more image
details. Figure 8 shows the generated local maxima mapping on the
accumulation array for selecting the area of interest which are the
possible area of the eyes in the image. Figure 9 shows the 3-D view
of the accumulation array after local maximum filtering. Figures 10
and 11 are the results from the Circular Hough Transform to detect
the circular shape of the eye pair in the image. Finally, Figure 12
shows the eye pair in the face image which was marked with a cross
(+) at the centre of the circle. From Figure 11, the 3-D view of
the accumulation array shows three points of the local maximum. Two
points for the eye regions and one point onthe hair of the person.
Since the Circular Hough Transform is used to detect the region of
the eye pair, only circular shape is detected as in Figure 12. As
the hair point is not circular in shape, the system did not detect
this point as the eye region on the face image.
The evaluation on the performance of the proposed algorithm is
carried out on 50 face images with opened eyes. Figure 13 (a) and
(b) are some examples of the face images for which the proposed
algorithm correctly detect the eye pairwhile Figure 14 (a) and (b)
show some example of which the proposed algorithm failed to detect
the eye pair. Since Circular Hough Transform detects circular
shape, the algorithm detects another circular shape on the face
image due to the circular shape of the nostril.
(a) (b)Fig.13. Detected eye pair
This happens because the face is tilted up and the nostrils are
exposed that cause the algorithm to wrongly detects the nostrils as
the eye region due to its circular shape Other factor could be due
to illumination since the circular white spot on the noise of
Figure 14 (b) was mistaken as the eye region.
(a) (b)Fig. 14. Wrong Detection of the eye pair
IV. CONCLUSION
The success rate of the eyes being detected is about 86% that
equals to 43 eye pair detected from 50 face images. The filter used
could not totally eliminate the effect of illumination variations
for all the images tested which causes the false detection of the
eyes. Perhaps using a better filter that able to
rectify this problem can increase the number of eye pair being
detected.Other then that, the combination of other techniquescan be
considered to eliminate the unwanted details on the face. However
the Circular Hough Transform is a relevant algorithm to be
considered in the eye detection process due to.
REFERENCES
[1] F. R. Brunelli and T. Poggio, Face recognition features
versus templates , Pattern IEEE Transaction Anal.ysis Machine
Intelligent15 (10) (1993) 10421052.
[2] Carlos Morimoto et al, Real-Time Detection of Eyes and
Faces,
http:www.cs.ucsb.edu/conferences/PUI/PUIworkshop98/papers/Morimoto.pdf,
1998.
[3] A.L. Yuille, P.W. Hallinan and D.S. Cohen, Feature
extraction from faces using deformable templates , International.
Journal,Computer. Vision 8 (2) (1992) 99111.
[4] X. Xie, R. Sudhakar and H. Zhuang, On improving eye feature
extraction using deformable templates, Pattern Recognition. 27
(1994) 791799.
[5] K.M. Lam and H. Yan, Locating and extracting the eye in
human face images, Pattern Recognition. 29 (1996) 771779.
[6] M.Nixon, Eye Spacing measurement for facial recognition,
Proceeding of the Society of Photo-Optical Instrument Engineers,
1985.
[7] A. Pentland, B. Moghaddam and T. Starner, View-based and
modular eigenspaces for face recognition, Proceeding. IEEE Conf. on
Computer Vision and Pattern Recognition (CVPR94), Seattle, WA,
1994
[8] W. min Huang and R. Mariani, Face detection and precise eyes
location, Proc. Int. Conf. on Pattern Recognition (ICPR00),
2000.
[9] J. Huang, H. Wechsler, Eye detection using optimal wavelet
packets and radial basis functions (rbfs), International Journal of
Pattern Recognition. Artificial Intelligent 13 (7) (1999)
10091025.
[10] G.C. Feng and P.C. Yuen, Variance projection function and
its application to eye detection for human face recognition,
International Journal of Computer. Vision. 19 (1998) 899906.
[11] G.C. Feng and P.C. Yuen, Multi-cues eye detection on gray
intensity image, Pattern Recognition. 34 (2001) 10331046.
[12] S. Kawato and J. Ohya, Real-time detection of nodding and
head-shaking by directly detecting and tracking the between-eyes,
Proceeding 4th IEEE Int. Conf. on Automatic Face and Gesture
Recognition, 2000, pp. 4045.
[13] Y. Tian, T. Kanade and J.F. Cohn, Dual-state parametric eye
tracking, Proceeding 4th IEEE Int.ernational Conference on
Automatic Face and Gesture Recognition, 2000.
[14] S.A. Sirohey and A. Rosenfeld, Eye detection in a face
image using linear and nonlinear filter, Pattern Recognition 34
(2001) 13671391.
[15] G. Chow and X. Li, Towards a system for automatic facial
feature detection, Pattern Recognition 26 (1993) 17391755.
[16] H.A. Rowley, S. Baluja and T. Kanade, Neural network-based
face detection, IEEE Transaction Pattern Analysis Machine
Intelligent.20 (1) (1998), 2338.
[17] K.K. Sung and T. Poggio, Example-based learning for view
base human face detection, IEEE Trans, Pattern Analysis Machine.
Inteligent. 20 (1) (1998) 3951.
[18] P.C. Yuen, G.C. Feng and J.P. Zhou, A contour detection
method: initialization and contour model, Pattern Recognition
Letter 20 (2) (1999) 141148.
[19] K. Sobattka and I. Pitas. "A Novel Method for Automatic
Face Segmentation. Facial Feature Extraction and Tracking," Signal
Processing:Image Communicarion, 12(3). 1998, 263-281.
[20] Huang, T.S., G.J. Yang, and G.Y. Tang, A Fast
Two-Dimensional Median Filtering Algorithm. IEEE Transactions on
Acoustics, Speech, and SignalProcessing, 1979. ASSP-27: p.
13-18.
[21] R. Gonzalez and R. Woods, Digital Image Processing;
Prentice Hall 2002.
[22] http://agingmind.cns.uiuc.edu/facedb/
242