Content uploaded by Gang Lu
Author content
All content in this area was uploaded by Gang Lu on Jan 25, 2019
Content may be subject to copyright.
A New Edge Detection Algorithm for
Flame Image Processing
Tian Qiu, Yong Yan* FIEEE, Gang Lu SMIEEE
Instrumentation, Control and Embedded Systems Research Group
School of Engineering and Digital Arts, University of Kent, Canterbury, Kent CT2 7NT, UK
t.qiu@kent.ac.uk, y.yan@kent.ac.uk, g.lu@kent.ac.uk
Abstract— Digital image processing is playing an increasingly
important part in imaging based flame monitoring systems. A
crucial step in flame image processing is to detect the flame edge,
a dividing boundary between the area where there is thermo-
chemical reaction and those without. The determination of flame
edges is a precursor to flame image processing and measurement
of flame parameters. Several known edge detection methods have
been tested to identify flame edges but the results achieved are
disappointing. As a result of recent work in feature analysis of
flame images, a novel flame edge detection method has been
developed, which can detect the flame edges effectively and
efficiently.
Keywords- image processing; combustion; flame image; edge
detection
I. INTRODUCTION
To meet the stringent standards on combustion efficiency
and pollutant emissions, flame monitoring with a quantitative
means is becoming increasingly important in foil-fuel
combustion systems, particularly in power generation plants [1].
This has led to a wave of research in advanced flame imaging
technologies [2, 3], not only in the power generation industry,
but also in fire safety engineering [4].
Edge detection is one of the important steps in flame image
processing. There are several reasons why it is necessary to
identify flame edges. First, flame edges determine the region of
a flame. This is especially important in the 3-D reconstruction
of the flame as without clearly defined edges, the flame cannot
correctly be reconstructed. Second, the use of flame edges can
reduce the amount of processing data and filter out unwanted
information, such as background noise within the image. In
other words, the edge detection can preserve the important
structural properties of the flame and meanwhile shorten the
processing time. Third, the edge detection can be used to
segment a group of the flames. This is particularly important
for the multiple flame monitoring in practical furnaces where a
multi-burner system is normally used. Finally but most
importantly, the flame edges form a basis for the quantitative
determination of a range of flame characteristic parameters.
A number of methods have been reported for extracting
flame edges for the geometric characterization of the flame [5,
6], and for the determination of fire location [7, 8]. A three-
dimensional (3-D) flame monitoring systems has also been
reported where three identical color CCD cameras together
with optical transmission units were used to concurrently
capture three images of a flame from different perspectives [5].
The outer edges of the flame images were detected and the
information obtained were used to build the geometric model
of the flame and thus to determine the size, volume and
location of the flame.
In this research several edge detection methods are
examined to assess their effectiveness for flame edge
identification. Despite many parameters are delicately adjusted
in the use of these methods, flame edges still cannot be
identified. It is therefore desirable to develop a dedicated edge
detection method for flame image processing. As a result of
recent work in characteristic analysis of flame images, a novel
computing algorithm is proposed where some of the unique
features of a flame image such as the number of main objects
and strong comparative brightness are used to identify flame
edges. Adopting these features, one can detect the coarse and
superfluous edges in a flame image and then identify the flame
main edges and remove the unrelated edges.
II. BASIC METHODS OF EDGE DETECTION AND THEIR
APPLICATIONS TO FLAME IMAGES
A typical edge in an image might for instance be the border
between blocks of different colors. The basic edge detection
method is to determine the level of variance between different
pixels by executing edge-detection operators for each small
matrix area. The edge-detection operator is calculated by
forming a matrix centered on a pixel chosen as the center of the
matrix area. If the value of this matrix area is above a given
threshold, then the middle pixel is classified as an edge.
Mathematically, the edges are represented by first- and
second-order derivatives. The first-order derivative (i.e.,
gradient) of a 2D function, f(x, y), is defined as a vector [9]:
， (1)
where Gx and Gy are the gradients in the x and y coordinates,
respectively. The magnitude of the vector is given by:
. (2)
The angle α, at which the maximum rate of change occurs, is:
. (3)
* Corresponding author
978-1-4244-7934-4/11/$26.00 ©2011 IEEE
281
All the gradient-based algorithms have kernel operators that
calculate the strength of the slope in directions which are
orthogonal to each other, commonly vertical and horizontal.
Examples of gradient-based edge detection operators include
Roberts, Prewitt, and Sobel operators. Then, contributions of
the different components of the slopes are combined to give the
total value of the edge strength. As an improved method using
the Sobel operator, the Canny edge detection algorithm is
known as an optimal edge detector [10]. In the present research,
these common edge detection methods have been attempted
with adequate parameters to process typical flame images.
Figure 1 shows example results obtained along with the
original image. As can be seen, those common edge detection
methods can only identify a part of the flame edges, or wrongly
identified small edges that are obviously not the edges of the
main flame.
(a) The original image (b) Matlab’s edge function (c) Canny method
(d) Prewitt metho
d
(e) Sobel metho
d
(f) Roberts metho
d
Figure 1. Representive results using the common edge detection methods.
III. A NOVEL EDGE DETECTION ALGORITHM
FOR FLAME IMAGE PROCESSING
Since flame images are a special class of images, some of
the unique features of a flame may be used to identify flame
edges. There are some differences between flame images and
other general images: flames are the main objects in the images;
the brightness of the flame is generally much higher than the
other objects while the background is comparatively dark;
furthermore, there is generally only one flame or one group of
flames in one image. The expected flame edge should be one
and only one clear, continuous, and uninterrupted edge.
Adopting these features, we have developed a novel flame edge
detection algorithm. The algorithm can be divided into several
logical steps as follows:
Step 1: Adjust the gray level of a flame image according to
its statistical distribution.
Step 2: Smooth the image to eliminate noise.
Step 3: Use the Sobel operator to find basic edges. This is
achieved by finding the gradients of all the pixels in the image
so as to highlight the regions with high gray level contrast at
their edges. The algorithm then tracks along those regions and
suppresses any pixels that are not at the peaks of gradients. If
the magnitude of the gradient is above the high threshold TH, it
is deemed an edge. And if the magnitude is between the two
thresholds, i.e., high threshold TH and low threshold TL, then it
is set to zero unless there is a path from this pixel to a pixel
with a gradient above TL.
The Sobel operator performs a two dimensional spatial
gradient measurement on an image. Then the approximate
absolute gradient magnitude (edge strength) at each point can
be found. It uses a pair of 3x3 convolution masks, one
estimating the gradient in the x-direction (columns) and
another estimating the gradient in the y-direction (rows). The
Sobel operator is expressed as follows [11]:
, (4)
. (5)
Step 4: Adjust TH and TL to get a better result. This is done
by giving the first pair of TH and TL initial values according to
the apriori results of similar flame images, and then adjusting
the values for a better result. The ‘better’ result is assessed by
that how many edges there are: the more edge pixels detected
in the edge image, the better the parameters are. Another
threshold TE is also set to restrict the total number of edges, i.e.,
if the number of edge pixels exceeds TE, the automatic
adjustment will be terminated. Until now, a preliminary image
with edges identified is obtained from the original flame image.
It is designated as a Preliminary Edge Image (PEI).
Step 5: Remove unrelated edges in the PEI through the
following steps;
5a) Select any edge point in the PEI, remove that point
from the PEI, allocate a new temporary edge image and plot
the point onto the temporary edge image.
5b) Use the selected point as the center and search in a 3×3
area. Store the location of all the neighboring pixels if they are
edge pixels. In eight neighboring pixels, operations are taken
for the following three different cases,
─ If there is no neighboring pixel, the selected point is then
an isolated point, and should be removed from the PEI.
Terminate the search and go to Step 5d.
─ If there is one neighboring pixel, the selected point is an
end point. It should then be removed from the PEI, plotted
onto the temporary edge image, and added into the endpoint
list. Start the new search from the found neighbor and go to
Step 5c.
─ If there are more than two neighboring pixels, then the
selected point is a normal transition point in an edge line or
an intersection with more than three bifurcations. Set one of
282
the neighboring points as the new search center and start a
search. Store the other positions as unchecked conjunction
points, and then go to Step 5b.
Figure 2 illustrates how the tracing step moves forward if
the old search center is replaced by a new search center. In the
pixels of the left image in Figure 2, pixel ‘5’ is the center
selected. Suppose an edge point at pixel ‘9’ is found, then
remove pixel ‘5’ from the PEI to temporary edge image, and
pixel ‘9’ will be the new search center. In this way, the search
moves forward pixel by pixel.
Figure 2. Illustration of movement of the edge search.
5c) Check the conjunction points. If all the conjunction
points have been searched as a center, one temporary edge
image is then completed. Compute the lengths of any two end
points in the temporary edge image and pick out the longest
one. Then go to Step 5d.
5d) If all the pixels in the PEI are moved to the temporary
edge image, then go to Step 6.
Step 6: Plot the pixels of the longest edge in the final edge
image which should have the same size as the original image.
The whole process is then complete.
The flow chart of the whole process is given in Figure 3.
IV. EXPERIMENTAL RESULTS AND ANALYSIS
After implementing the algorithm as described in Section
III, hundreds of flame images were processed using the
algorithm so as to evaluate its effectiveness. The images used
were taken for propane Bunsen flames burning in open air. The
results have shown that clear flame edges have successfully
been identified in all the flame images. Figure 4 shows typical
processed flame images with edges identified. In comparison
with the test results presented in Figure 1, it can be clearly
observed that the developed algorithm can successfully detect
the clear edges of the flame and disregard unrelated small
edges, which the common edge detection methods cannot do. It
makes much easier to distinguish the flame region from the
background. The algorithm can also be used to extract the
edges of complex flames such as turbulent diffusion flames or
flames of pool fires [12]. The clearly defined flame edges will
form a basis for subsequent processing of the flame images
such as flame size computation, flame background removal,
and determination of other flame parameters [3].
Although the algorithm appears to be long, it is not tedious
in computation. Tested on a desktop computer with 2.66GHz
Intel® Quad CPU, it can detect 180 flame edges of 141×161
pixels flame images in one minute.
Figure 3. Flow chart of the flame edge detection algorithm.
283
(a)
(b)
(c)
(d)
(e)
(
f
)
Figure 4. Some of the flame edge detection results [Left column: original
images; right column: images with identified edges. (a) Diffusion pronane-
flame, (c) Partially-premixed pronane-flame, (e) Small scale pool fire[13]].
V. CONCLUSIONS
After flame characteristics are analyzed, a new flame edge
detection method has been developed and evaluated in
comparison with conventional methods. Experimental results
have demonstrated that the algorithm developed is effective in
identifying the edges of complex and irregular flames in noisy
images. This reasonably fast and convenient flame edge
detection method lays a good foundation for subsequent
quantification of flame parameters and 3D flame reconstruction
and visualization.
ACKNOWLEDGMENTS
Acknowledgments are made to the RCUK’s Energy
Programme for grants (EP/G062153/1 and EP/G063214/1) in
aid of this research, but the views expressed are those of the
authors, and not necessarily those of the RCUK. The Energy
Programme is an RCUK cross-council initiative led by the
EPSRC and contributed to by ESRC, NERC, BBSRC and
STFC.
REFERENCES
[1] D. Roddy, Advanced Power Plant Materials, Design and Technology,
Woodhead Publishing, ISBN 1 84569 515 1, 2010.
[2] G. Lu, Y. Yan, and M. Colechin, “A digital imaging based multi-
functional flame monitoring system”, IEEE Trans. Instrum. Meas., Vol.
53, No. 4, pp. 1152-1158, 2004.
[3] Y. Yan, T. Qiu, G. Lu, M. M. Hossain and G. Gilabertet, “Recent
advances in 3D flame tomography”, Proceedings of the 6th World
Congress on Industrial Process Tomography, Beijing, China, pp. 1530-
1539, September, 2010.
[4] D. S. Huang, L. Heutte, and M. Loog, “Real-time fire detection using
camera sequence image in tunnel environment”, ICIC 2007, LNCS
4681, pp. 1209–1220, 2007.
[5] G. Lu, G. Gilabert and Y. Yan, “Vision based monitoring and
characterisation of combustion flames”, Journal of Physics: Conference
Series 15, pp. 194–200, 2005.
[6] H. C. Bheemul, G. Lu, and Y. Yan, “Three-dimensional visualization
and quantitative characterization of gaseous flames,” Meas. Sci.
Technol.,vol. 13, No. 10, pp. 1643–1650, 2002.
[7] B. C. Ko, K. H. Cheong, and J. Y. Nam, “Fire detection based on vision
sensor and support vector machines ”, Fire Safety Journal, Vol. 44, pp.
322–329, 2009.
[8] Q. Jiang, and Q. Wang, “Large space fire image processing of improving
canny edge detector based on adaptive smoothing”, 2010 International
Conference on Innovative Computing and Communication and 2010
Asia-Pacific Conference on Information Technology and Ocean
Engineering, 2010.
[9] R. C. Gonzalez and R. EWoods, Digital Image Processing, Second
Edition, Prentice Hall, 2002.
[10] T. Steinhaus, S. Welch, R. Carvel, and J. L. Torero, “Large-scale pool
fires”, Thermal Science Journal, Vol. 11, No. 3, special on fire, 2007.
[11] J. Canny, “A computational approach to edge detection”, IEEE
Transactions on PAMI, Vol. 8, No. 6, pp. 679–698, 1986.
[12] K. R. Castleman, Digital Image Processing, Prentice Hall, 1995.
[13] SWRI, “Small-scale liquid pool fire characterization”, Available at:
http://www.swri.org/4org/d01/fire/firetech/about.htm.
284
