IMPROVING NEURAL NETWORK-BASED FACE DETECTION

advertisement
Shoulder Detection Using GRELD Threshold Method*
Ming Zhang*
Jeffrey C. Bailey
Department of Physics, Computer Science and Engineering, Christopher Newport University
1 University Place, Newport News, VA, USA
* This research is supported by Christopher
Newport University Applied Research Center.
Abstract
Actual detection of a face or faces within an
image outside of controlled laboratory
conditions remains a challenge due to
environmental factors such as lighting,
orientation, background, occlusions, and altered
appearances. This paper developed a new
method called Golden Ratio and Edge Line
Detecting (GRELD) threshold method. GRELD
threshold
algorithm
is
provided
too.
Experimental result shows that GRELD
threshold method is 22.4% better than basic
threshold method for shoulder detection. With
the better accuracy rate of shoulder detection,
then the face detection success rate might have
been improved.
Keywords
Shoulder Detection, GRELD,
Method, Face Detection
Threshold
1. Introduction
Human face detection in video or still images
has long been an area of interest for many
researchers.
Automatic face detection and
recognition in images has numerous applications
in computer systems today, including computer
vision
for
human-computer
interaction,
recognition for criminal and other security
systems, identification, and access control. The
role of face detectors in face recognition systems
is critical to their success. It is necessary to
understand the difference between face detection
and face recognition. Yang et al [1] makes the
distinction.
Face detection is: “Given an
arbitrary image, the goal of face detection is to
determine whether or not there is a face in the
image and, if present, return the image location
and extent of each face [1].” Face recognition
involves comparing a face that has been located
in an image with a database of faces with the
intent of finding a match. The greatest challenge
facing computer scientists is face detection.
Most face detection systems use either template
matching methods or appearance-based methods
[3]. Other methods, such as knowledge-based or
feature-based, are more effective in face
localization problems, which are simplified
detection problems [1]. Zhang et al [2] uses an
artificial neural network group-based adaptive
tolerance (GAT) tree model and shows that it is
able to solve complex pattern recognition
problems that involve non-continuous, nonsmooth decision functions.
A number of edge detection algorithms have been
developed and are used in pattern recognition
applications. One of the more successful edge
detectors is Canny’s. Heath et al [4], which
prescribed a method of quantifying subjective
performance assessments of different edge detectors,
calls Canny’s algorithm the “modern standard”.
Because it is the one new detectors are most often
compared to. Some face detectors rely on
component-based detection to detect a face by
locating individual facial components and
considering the results to determine whether a face
has been detected.
When examining various face detection systems,
one will notice there are two types of errors
made by theses systems: false positives and false
negatives. False positives occur when a face
detection system incorrectly detects and locates
a face in an image where a face does not exist.
On the other hand, false negatives result from
the system’s failure to detect a face in an image
when it does exist. By considering the failures of
face detection systems, whether false positives
or false negatives, it is interesting to note that
had the presence of human shoulder been
detected (in the case of false negatives) or had
the lack of human shoulders been detected (in
the case of false positives), then the detection
success rate might have been improved.
Likewise, since human shoulders, in many cases,
form distinct edges, the use of an edge detection
algorithm to detect them is a natural choice. This
paper developed a new method called Golden
Ratio and Edge Line Detecting (GRELD)
threshold method. Using the GRELD threshold
method to produce a better accuracy rate of
shoulder detection, then the face detection
success rate might have been improved.
2. Moving Shoulder Detector(MSD)
The Moving Shoulder Detector (MSD) has been
built based on a PC, using Visual Studio.net
C++, PixelSmart capture card, and Victor image
processing library. The structure of the MSD
can be seen in Figure 1.
Shoulder Output Display
Moving Shoulder Detection
EVS(i,j,t)
DOF(i,j,t)
GRELD
Threshold
Method
First-Order and
Second-Order
Regime Motion
Detection
In order to obtain input moving face image,
MSD will be fed moving facial images from
CCD camera, using the appropriate signal
formats. This moving face image vector, MI(i,
j, t), will be input to the GRELD threshold
subsystem for shoulder edge vector. It will also
be sent to the first/second–order motion
model(s) to determine motion direction
information.
 MI(i, j, t): moving image vector which can
be used to represent a human face
 DOF(i, j, t): detected direction vector of
face moving
 EVS(i, j, t): edge vector of shoulder
moving
The input moving face image vector will be used
as input to the first-order linear regime, while
the power of the vector will be used for secondorder non-linear rectifying regime motion
detection. Now while the first-order linear
regime model may fail to detect motion
direction, the second-order non-linear rectifying
regime model will do so [6]. The detected face
motion direction vector, DOF(i, j, k), will be
input in turn to Moving Shoulder Detection
Threshold is a basic method for edge detection.
But simple threshold always cannot provide
good edge detection result. A new method,
called GRELD threshold method, has been
developed. The output of GRELD threshold
method is EVS(i,j,t) - edge vector of shoulder
moving. Finally, DOF(i, j, k) and EVS(i,j,t)
will be used to detect the moving shoulder.
Details of GRELD threshold method can be
found in next section.
Power
3. GRELD Threshold Method
Image Preprocessing and Normalization
Input Moving Face Image Vector MI (i, j, t)
CCD Camera
Fig. 1. MSD: A Moving Shoulder Detector
The GRELD threshold method can be described
as follows:
Step1:
Input Moving Face Image Vector MI (i,j,t);
Step2:
Choose the threshold value Tn(n=1,2, ….,n)
Tn=0.618*R (n= 1,2, …., n)+ Low Boundary of T
Rn : the nth threshold value range
Step3:
if ( MI(i,j,t) > Tn ) MI(i,j,t) = Va
else
MI(i,j,t) = Vb
Va: the highest value of pixel in the image
Vb: the lowest value of pixel in the image
Step 4:
MI(i,j,t) =MI(i,j,t)/2 + MI(i,j+1,t)/2
Step5:
if ( ( MI(i,j,t) > Tu ) && ( MI(i,j,t) < Tl ) )
MI(i,j,t) = Vb
else
MI(i,j,t) = Va
Step 6:
If shoulder has been detected, stop.
Otherwise go to Step 2.
The formula of (Tn =0.618*Rn + Low Boundary
of T) is based on the Golden Ratio method [5].
Using the number 0.618, it will guarantee that
the best value will be found under the minimum
steps. To detect the shoulder and draw the
shoulder lines, Step 3 and Step 4 have been
added. Otherwise, no shoulder line could be
detected. The nth threshold value range Rn will
depend on the MI(i,j,t) and Golden Ration
method [5]. The following Table 1 is an
example to demo how to choose the Tn and Rn .
Table1 Example of Tn and Rn Values
Loop Rn, Value Calculation
Tn
1
R1 255
255 - 0 =255
T1 158
0.618 * 255 + 0 ≈ 158
2
R2 97
255 - 158 =97
T2 219
0.618 * 97 + 158 ≈ 219
3
R3 61
219 - 158 =61
T3 200
0.618 * 61 + 181≈ 200
4
R4 19
219 - 200 =19
T4 213
0.618 *19 + 200≈ 213
5
R5 23
213 - 200 =13
T5 209
0.618 * 13 + 200≈ 209
6
R6 9
209 - 200=9
T6 206
0.618 * 9 + 200≈ 206
7
R7 6
206 – 200 = 6
T7 204
0.681 * 6 + 200 ≈ 204
8
R8 4
204 -200 = 4
T8 203
0.681 * 4 + 200 ≈ 203
9
R9 3
203-200 = 3
T9 202
0.681*3 +200 =202
10
R10 2
202 – 200 = 2
T10 201
0.681 * 2+200 = 201
4. Experiment Results
Figure 2A and 3A used basic threshold method
for shoulder detecting. The basic threshold
method uses following formula:
If ( ( MI(i,j,t) > Tl ) && (MI(i,j,t) <Tu ) )
MI(i,j,t) = Va
else
MI(i,j,t) = Vb
Tl : Low threshold value
Tu: Up threshold value
Va: the highest value of pixel in the image
Vb: the lowest value of pixel in the image
Figure 2B and 3B used GRELD threshold
method for shoulder detecting. Figure 2A and
2B have no noise added. In Figure 3A and 3B
5% random noise has been added. It is easy to
find, based on the Figure 2A, 2B, 3A, and 3B,
GRELD threshold is better than the basic
threshold method.
Table 2 provided the comparison of two
methods, basic threshold method and GRELD
threshold method. Experimental results told us,
if we add 0% to 20% random noise, the GRELD
average SPD (Shoulder Points Detected)/ SP
(Shoulder Points) is 46.4% better than basic
threshold method. The GRELD average of SNP
(Shoulder Noise Points)/SP (Shoulder Points) is
22.4% better than basic threshold method.
Fig. 2A Basic Threshold Method and 0% Noise
Fig. 2B GRELD Threshold Method and 0%
Noise
Fig. 3A Basic Threshold Method and 5% Noise
the better accuracy rate of shoulder detection,
then the face detection success rate might have
been improved. As a next step, we will study the
shoulder detecting of different background. We
already noticed, different background could
make the shoulder detecting more difficulty.
Automatically choosing and adaptively selecting
the best threshold values are our future research
area too. Finally, we will build a moving face
detecting system with very high accuracy rate.
Acknowledgment
The authors would like to thank Prof. A. Martin
Buoncristiani for his great support. The authors
would also like to express their thanks to Prof.
David Hibler and Dr. Antonio Siochi for their
invaluable help.
Fig. 3B GRELD Threshold Method and 5%
Noise
Table 2 Threshold Methods Comparison
Method
Noise SPD/SP SNP/SP
Threshold
0%
80
5%
Threshold
5%
30%
7%
Threshold
10%
55%
10%
Threshold
15%
30%
50%
Threshold
20%
50%
90%
Average
49%
32.4%
Method
Noise SPD/SP
GRELD
0%
100%
GRELD
5%
100%
GRELD
10%
95%
GRELD
15%
92%
GRELD
20%
90%
Average
95.4%
SPD: Shoulder Points Detected
SNP: Shoulder Noise Points;
SP: Shoulder Points
SNP/SP
0%
0%
0%
20%
30%
10%
References
[1]
[2]
[3]
[4]
5. Summary
This paper developed a new method called
Golden Ratio and Edge Line Detecting
(GRELD) threshold method. GRELD threshold
algorithm is provided. Shoulder detection has
been studied by using GRELD threshold
method. Experimental result shows that GRELD
threshold method is 22.4% better than basic
threshold method for shoulder detection. With
[5]
[6]
M. H. Yang, D. J. Kriegman, and N.
Ahuja. Detecting Faces in Images: A
Survey. IEEE Trans. Pattern Analysis
and Machine Intelligence, 24(1): 34–58.
2002.
Ming Zhang, and John Fulcher. Face
Recognition Using Artificial Neural
Network
Group-Based
Adaptive
Tolerance (GAT) Trees. IEEE Trans.
Neural Networks. 7(3): 555–567. 1996.
Ming Zhang, and John Fulcher. Face
perspective
understanding
using
artificial neural network group-based
tree. Proceedings of International
Conference on Image Processing,
Lausanne, Switzerland, vol. III,
September 16-19. pp.475-478. 1996.
M. Heath, S. Sarkar, T. Sanocki, and
K.W. Bowyer. A Robust Visual Method
for Assessing the Relative Performance
of Edge-Detection Algorithms. IEEE
Transactions on Pattern Analysis and
Machine Intelligence. 19(12):13381359. 1997.
Mario Livio. The Golden Ratio.
Broadway Book. pp. 1-11. 2002.
Chubb, C., & Sperling, G.. Driftbalanced random stimuli: a general basis
for studying nonFourier motion
perception. J. Optical Society of
America. 5:1986-2007. 1988.
Download