Multi-Emissivity Setting in Thermal Imaging Based on Visible-Light Image Segmentation

TELKOMNIKA Indonesian Journal of Electrical Engineering
Vol.12, No.1, January 2014, pp. 245 ~ 253
DOI: http://dx.doi.org/10.11591/telkomnika.v12i1.3913  245
Received June 21, 2013; Revised August 9, 2013; Accepted September 14, 2013
Multi-Emissivity Setting in Thermal Imaging Based on
Visible-Light Image Segmentation
Ding De-hong1,a
, Fang Kui*1,b
, Yu He-xiang1
, Qian Ke-jun1
Chen Yi-neng1
, Cui Dai-jun2
1
College of Resource & Environment, Hunan Agricultural University, Changsha, PR China
2
Department of R & D, YADO Monitoring Technology Co. Ltd., Zhuhai, PR China
*Corresponding author, email: fk@hunau.netb
, dingdehong@qq.com
a
Abstract
Emissivity is an accuracy influencing factor during infrared temperature measurement which
focusing on regular geometry in laboratory under the condition of single emissivity. But in practical
application, the target is often irregular with multi-emissivity,, and if following "idealized" method in
laboratory, it will lead to inevitable error. This paper presents a method for a complex target object with the
collection of multiple emissivities in infrared image after measurement. Both visible and infrared images
were collected in the same field of view at the same time using binocular video to segment target
regionally through visible image. The emission rate in corresponding region was set based on regional
growing algorithm. Heat conduction equation was used as a reference to smooth the boundary area. After
testing image evaluation parameters accordingly, results obtained via this infrared temperature
measurement method are closer to the true value and precise compared with conventional ones judged
from objective measurements.
Keywords: Multiple emissivity setting, edge extraction, region growing segmentation, boundary smooth
transition base on equation of heat conduction
Copyright © 2014 Institute of Advanced Engineering and Science. All rights reserved.
1. Introduction
There are mainly two blackbody radiation theories, Stefan - Boltzmann law and Planck
radiation law which describes the relationship between the radiation rate and frequency of the
electromagnetic radiation emitted from a blackbody, and the spectrum curves of blackbody
radiation under different temperatures do not intersect. Stefan - Boltzmann law is as follows.
TW
4
 (1)
All infrared temperature measuring devices are based on the principle that the surface
temperature of objects can be accurately determined by measuring the infrared energy radiated
by themselves. Blackbody is an idealized radiating body with a surface-emitting rate of 1. The
radiation amount of actual objects existing in nature world depends on not only radiation
wavelength and temperature of objects, but also type of material constituting object, surface
roughness, physical and chemical structure, material thickness, preparation method, thermal
process, surface state and environmental conditions etc [1]. For the application of blackbody
radiation law to all actual objects, emissivity  has been introduced as a coefficient related to
material properties and surface state. This coefficient represents the approximation degree
between heat radiation of actual objects and that of blackbody, and
(0,1) 
. According to the
law of radiation, the infrared radiation characteristics of any objects will be acquired as long as
the emissivity of material is obtained.
The commonly used methods to determine the emission rate are Fourier transform
infrared spectroscopy, double-reference method, dual-temperature method and dual
background method, but they only target regular geometry and the results are single emissivity.
Actually, the tested target is often of irregular geometry with the presence of multiple
emissivities. If we follow the "ideal" laboratory method, a pre-set single emissivity will be
necessary, which will lead to an inevitable error. Wang Pinqing [2] also pointed out that
inaccurate emissivity has a significant impact on the tone display in geological infrared heat

 ISSN: 2302-4046
TELKOMNIKA Vol. 12, No. 1, January 2014: 245 – 253
246
map, but did not propose solutions. American scholars, Watson and Kahle [3-4], have
recognized the influence of emissivity variation of ground objects and geological body on the
calculation of surface temperature, considering that imaging using two thermal infrared bands
(8-14 μm and 10.5-11.5 μm) can regulate the error of temperature calculation due to the
changes of emissivity within the range 1%, the conclusion of which is in accordance with
document [5]: for blackbody at 150 ℃, when measurement error of emissivity is 5%, the
theoretical calculation error of measured temperature with thermal imager is up to about 1%.
It is also pointed out that [2] the actual surface temperature of different objects or
geological bodies with the same tone (i.e. same radiation energy) in thermal infrared scanning
image may vary considerably. For instance, water body at 27 ℃, quartz at 64 ℃ and marble at
32 ℃ may have the same radiation energy, i.e. same radiation temperature [6]. One of the
methods infrared camouflage [7]—Making un-isothermal objects look isothermal, can be
regarded as a example of reverse application involves emissivity. Generally, an emissivity
difference of 0.01 among objects might result in an error of about 1 ℃ of radiation temperature.
A method targeting complex object has been proposed in our study based on visible image
segmentation technology, by setting multi-emissivity in infrared heat image afterwards, and can
carry out infrared radiation thermometry more accurately.
2. Image Acquisition
2.1. Dual-channel Image Capture
Analysis and studies have shown that infrared thermal images (IR) and visible images
(VI) are different in the following situations [8]: imaging of IR is based on the radiation caused by
the temperature of object (passive imaging), while imaging of VI on the reflectivity of object
(positive imaging); spatial resolution of IR is generally lower than that of VI. Grayscale of VI is
more structured compared to IR’s; texture details on the surface of scene can be reflected on VI
but not IR. Edge characteristics in VI are evident with relatively steep edge while that of IR
relatively smooth. Simultaneously, there is missing and offset phenomenon occurring at the
edge of IR compared to VI for the same scene. Owing to the existence of more low-frequency
components of IR, the correlation length of IR is larger than that of VI for the same scene.
2.2. Experimental Conditions
An online monitoring infrared camera (FLIR movement, DAHENG processor) purchased
from Yado monitor company in Zhuhai City was used with a default emissivity of 0.95 and
output mode of color selected as “Fusion”. The indoor temperature of laboratory environment is
20 ℃, relative humidity 65%, and the room is without air convection with closed doors and
windows. The size of aluminum block as research object is 20 * 10 * 5 mm with matte surface,
and the intermediate was coated with paint into an irregular shape. After the block was placed in
room for 2 hours for heat exchanging balance, shooting was conducted with respect to the
same visual field. Figure 1 and 2 are VI and IR respectively under same visual field with a
resolution of 320 * 240.
Figure 1. Visible-light image Figure 2. Infrared thermograph
2.3. Issue Introduction
The aluminum block was maintained a consistent temperature as a whole by great
thermal conductivity. It is obvious to be observed visually that, compared with temperatures

TELKOMNIKA ISSN: 2302-4046 
Multi-emissivity setting in thermal imaging based on visible-light image segmentation (Fang Kui)
247
bands, the temperature of black paint amid the block is significantly higher than that of
surrounding empty area by 4.4 measured with Yado’s infrared camera, which means that the
emissivity of these two regions is different. In other words, the emissivity of one object to be
measured can not be simply fixed, getting rid of the method described in literature [5]. It is
presented in this paper a method that multi-emissivity is set according to regions and how to
shoot isosbestic infrared image for isothermal objects. Technology and methods involved are
described as follows.
3. Edge Definition Based on Visible Light Image Segmentation
Partition the visible image because of the distinctive edge characteristics of VI. Image
segmentation is separating image into different regions of specific meanings and disjoint, and
each region is of the consistency of specific area. A image g (x, y) segmentation is dividing
image into sub region g1, g1 and g3 meeting the following conditions:
1
( , ) ( , )
N
kK
g x y g x y
U I.e., all regions comprise the whole image.
kg
is a region in communication.
( , ) ( , )k jg x y g x y I
I.e., two arbitrary sub regions have nothing in common.
Region kg
is of uniformity to a certain extent, which refers to small difference or slow
changes of grayscale value between the pixels within the same region.
There are many methods for image segmentation, and not an individual image
segmentation algorithm but the combination of several methods are applied to achieve perfect
image segmentation effect [11-12]. Classic methods include threshold segmentation [9-10],
edge detection method [13], statistical segmentation method [14] and the method of combining
region and boundary information. In addition, there are many other methods and literatures in
the field of image segmentation, such as image segmentation based on partial differential
equation [15], image segmentation based on genetic algorithm [16], image segmentation
method based on wavelet analysis and transform [17-18], image segmentation method based
on mathematical morphology [19], image segmentation method based on fuzzy theory [20-21],
image segmentation method based on artificial neural network [22]. The black region in Figure 1
was segmented using Canny detector with the aid of MATLAB [23] in this paper. The
procedures are:
Image is smoothed by a Gaussian filter with specified standard deviation  to reduce
the noise.
Local gradient
2/122
][),( yx GGyxg 
and edge direction
)/arctan(),( xy GGyx 
are
calculated at each point; edge point is defined as the point of maximum local strength in
gradient direction.
Edge points defined in item (b) will result in ridge in gradient magnitude image. Then,
the algorithm tracks top of all the ridges and set all pixels not at the top of ridge to zero in order
to create a thin line in the output, which is well-known as non-maximum resisting treatment.
Pixels at the ridge are set using two threshold values 1T and 2T ( 1 2T T ): ridge pixel is called
strong edge pixel when the value is greater than T2 while called weak edge pixel when between
1T and 2T .
Finally, the algorithm integrates weak pixels connected by 8 pixels into strong pixels
and performs edge linking.
Canny edge detector is performed using the following
syntax:
[g, t]=edge (f, 'canny', T, sigma) (2)
whereT is a vector and [ 1, 2]T T T
. Sigma is the standard deviation of smoothing filter
with a default of 1.
Procedures of edge extraction from VI then superimposed with IR in Matlab are
described as follows:
rgb=im2double(imread('IR_1.jpg'));
IR=imread('VI.jpg');
I=rgb2grayscale(IR);

 ISSN: 2302-4046
248
K=imadjust(I,[20/255 1]); % contrast enhancement
Ir_edge=edge(K, 'canny',[0.04 0.50],1);
r=rgb(:,:,1).*~Ir_edge;
g=rgb(:,:,2).*~Ir_edge;
b=rgb(:,:,3).*~Ir_edge;
r(Ir_edge)=1; % edge set as white
g(Ir_edge)=1;
b(Ir_edge)=1;
result=cat(3,r,g,b);
Figure 3 is the extracted edge and Figure 4 the result of regional definition after
superposition of Figure 2 and 3.
Figure 3. Extraded edge Figure 4. Infrared thermograph after regional definition
4. Multi-Emissivity Setting According to Region
The following conclusions can be drawn from the comparison of Figure 2, for instance,
and equation 1: under ideal condition, W is a constant and T is inversely proportional to . If
actual emissivity is higher than the one set by instruments, the measured temperature will be
excessively high, and lower actual emissivity will lead to lower measured temperature, on the
other hand. High value turns higher and low value lower. A greater error of  will cause more
chromatic aberration. In other words, chromatic aberration of heat map is larger than the actual.
After setting emissivity according to region chromatic aberration should be reduced, and the
whole aluminum block in Figure 2 should be in the same uniform color under ideal
circumstances.
a) Determination of emissivity
Paint of a known emissivity was evenly sprayed on the surface of tested object, and
then adjusted the emissivity of infrared camera until the exposed and painted surface reached a
same or close temperature. Emissivity at this time is the correct one of the target [24].
b) Multi emissivity setting on the basis of region growing algorithm
Emissivity (Es) of the paint is 0.99 by looking up relevant table and that of non-paint part
Emissivity (Er) 0.3 determined by the above experimental methods. After the setting of painted
part emissivity with Es and non-paint with Er respectively region growing algorithm was used
with Matlab [23]:
Q= True; Absolute value of grayscale difference between pixels of seeds and a point
( , )x y T
Q= False; Other situations
The first step is to determine seed, which is conducted with mouse (command “getpts”
in Matlab) manually. Then, all the pixels corresponding to the following conditions were attached
to each seed to form growing area: (a) connected to seed in the form of 8, and (b) similar to the
seed. T is a designated threshold value, although this feature is based on grayscale-scale
difference and relatively unitary. A more complicated scheme is formulated in this research that
the corrected grayscale heat map is derived via reverse derivation according to the two different
emissivity (Es and Er) and original grayscale scale image, and then converted to pseudo-color

249
image. As a result, the paint part shifts to color blue and non-paint portion to color orange.
Owing to the varied fitting equations correlating grayscale with temperature of various detectors,
even the identical detector, parameters employed by fitting equation will change in different
situations, so the expatiatory quantitative calculation process is ignored here, and the results
are shown in Figure 5 and 6, both of which are superposed to form Figure 7. Chromatic
aberration of both regions is significantly reduced visually and the pseudo-color image tends to
consistency, but obvious boundary defects are observed, which is because pixels at the junction
belong to neither Figure 5 nor Figure 6.
Figure 5. Non-pain region Figure 6. Painted region Figure 7. Superimposition of two
regions
5. Smooth Transition of “Color Temperature” at Stitching Place using heat Conduction
Equation
Joseph · Fourier describes a general heat conduction model in his published work
“analytic thermology” in 1822:
t xxu ku
The x is spatial variable
[0,1] indicating the normalized length, and t is time variable,
where
k = .061644 subject to the boundary conditions:
(0, ) 0 (1, )x xu t u t 
and with the initial heat distribution given by:
( ,0) cos(2 )u x x
In this case, the left face (x=0) and the right face (x=1) are perfectly insulated. This
image shows how the heat redistributes, flowing from the warmer left edge to the cooler right
edge, then equalizing to a constant temperature throughout. This temperature happens to be
the average value of cos(2x) over [0,1], as one might expect.
The solution [25]:
sin(2) 2 2
2
1
( , ) cos( ) exp( )n
n
u x t A n x kn t 


  
Where:
1
2 20
4 sin(2)
2 cos(2 ) cos( ) ( 1)
4
n n
A x n x dx
n



  

A transient thermal conductivity curve is shown in Figure 8 at some time-point. The
curve becomes flatter as evolution and horizontal when reaching thermal equilibrium.
It is hard to make the "colorful temperature"(which weighs the temperature by color)of
painting area as the totally same as that of non-painting area, due to the error of the emissivity
or effect of precision. If the two areas are the same completely, the result seems to be not
reality. But they should be the same on the theory, which means finding a way to make the
colorful temperature on the boundary transit smoothly. The figure 8 provides the reference to
solve this question. The situation of figure 7 could imitate as the way of heat conduction: the
heat conduction has already reach some balance that there is no change on transition band
between high-temperature zone and the low-temperature zone. Specifying that T2 indicates the

 ISSN: 2302-4046
250
high-temperature in painting area and T1 represents the low-temperature of non-painting area.
As the result of that, temperature of transition area, from the intrinsic to extrinsic, should from T2
to T1 smoothly according to the heat conduction equation. Choosing when and which part of
curve, it is not sure. On the sake of the difference between T1 and T2 is small, we can use
direct- line instead of Cn-C1 in this paper.
Figure 8. Transient heat conduction curve
Figure 9. Heat map after smooth transition for
boundary
The procedure using Matlab to implement this algorithm is: expand the contour line to a
width of n pixels (the original is 1 pixel wide), and n is an odd number to exactly cover the
vacant points at splicing. n = 5 in this paper actually. Loop point-by-point along the original
contour and find line segments Li with a length of n at point i perpendicular to contour line (node
coordinates of Li from outside to inside: Ci1 ... Cin; and contour line at the midpoint: Cin/2 + 1).
It can be obtained from Figure 8 that TCi1 = T2 and TCin = T1. Then divide line segment Li into
n-1 equal parts to deduce temperatures of other points. After processing, the result is shown in
Figure 9 and the main pseudo-code is given by:
For i = 1: M // Loop point-by-point along the original contour
Solve line segment Li
TCi1 = T2, TCin = T1
For j = 2: n-1 // calculate color and temperature of n-2 points in the line Li
TCij = TCi (j-1) + (T2-T1)/(n-1)
Reversely calculate grayscale value Grayscale-i according to TCij
Calculate color Color-i at the point according to Grayscale-i
Set the color at Cij to Color-i
End;
End;
6. Experimental Results and Parameters Test
Referring to characteristic properties of texture statistics [8], evaluate the grayscale
components of Figure 2 and 8 by citing the following statistical characteristics:
Information entropy: Information entropy of image is an important indicator measuring
the information richness in images and the size of which reflects the average amount of
information contained in images, which is defined as




1
0
2log
L
i
ii PPH
，where Pi is the distributional probability of grayscale i in the range
of
[0,1, , 1]L L . Greater entropy means richer information of an image.
Standard deviation: It represents the discretion extent between image grayscale and
average grayscale, which is defined as

251
NM
FnmF
M
m
N
n
*
]),([
1
0
1
0
2
 






where F is the average grayscale of the whole image. A larger standard deviation will
lead to more dispersed image grayscale distribution and greater image contrast, and more
information will be extracted. On contrary, concentrated image grayscale distribution will result
in smaller contrast and less information.
Contrast: Contrast refers to the grayscale ratio of bright and dark parts on screen,
which is given by:
( , ) ( , ) ( , )c r i j r i j p i j  
( , )r i j i j 
grayscale difference between adjacent pixels;
( , )p i j the pixel
distributional probability of gray-scale difference of r between adjacent pixels.
Smoothness: Smoothness characteristics [26] is based on grayscale histogram h (n):
254
0
( 1) ( 1)
( ) | ( ) | | (0) (1) | | (255) (254) |
2i
h i h i
S h h i h h h h

  
     
This parameter can illustrate, to a certain extent, the distributional smoothness of
grayscale histogram. If the histogram distribution is uniform, the calculated characteristic value
will be small.
Third moment: The skewness of histogram can be measured by this parameter. If the
histogram is symmetrical, the value is 0. If it is positive, histogram skews right, and if negative,
skews left.
)()(
31
0
3 i
L
i
i zpmz



Consistency: When all grayscale values are equal, this parameter reaches acme and
decreases in other situations.
Tested by “Crop Image Analysis System Program”[27], it can be concluded from Table
1 that entropy, standard deviation, contrast and third moment are all smaller in Figure 9 than
those in Figure 2 except for consistency and smoothness, meaning that heat map demonstrates
greater uniformity, smoothness and consistency, namely heat map of aluminum block with the
emissivity set according to regions, almost a whole, and objective evaluation is consistent with
subjective evaluation.
Table 1. Comparison of parameters for evaluation
Parameters Figure 2 Figure 9 Variance
Temperature difference(℃) 4.4 0.7 ↓
Entropy (e) 6.28666 5.09708 ↓
Standard deviation(  ) 0.33481 0.25940 ↓
Contrast (c) 0.15223 0.11496 ↓
Smoothness S(h) 0.00831 0.04782 ↑
Third moment(∪3) 2.55723 0.18665 ↓
Consistency U 0.01669 0.04644 ↑
7. Conclusion
This paper presents a method using improved region growing algorithm with setting
distinct emissivity at different locations through edge extraction based on visual image and then
superimposing the edge on infrared heat map. With the aid of heat conduction equation, smooth
transition zone to make the heat hologram indicate real temperature more precisely.

 ISSN: 2302-4046
252
Comparison Figure 9 with Figure 2, painted and non-paint regions tend to be more consistent
after processing, confirmed with experimental measured parameters. This method is extremely
appropriate for the heat monitoring of combination and mixture, such as monitoring respiration
and heat release of different seeds under the same ambience for nurturing, and
camouflage/anti-camouflage. Further engineering conversion will be required for applying this
research results to engineering practice under complex geological circumstances.
Acknowledgements
This work was supported by National Natural Science Foundation of China (Grant No.
61101235), "Twelfth Five-Year" National Science and Technology Program (Grant No.
2011BAD21B03), Science and technology major projects of Hunan Province, China (Grant No.
2010FJ1006); Youth science fund projects of Hunan Agricultural University, China (Grant No.
11QN68).
References
[1] Ye yutang. Infrared and Low Light Level Technology. Beijing:National Defense Industrial Press. 2010:
22-25.
[2] Wang pinqing. The Effect of Emissivity on Thermal’s image and It's Interpretation. Remote sensing of
environment. 1988; 15: 207-212.
[3] Alan R, Gillespie, Anne B, Kahle. Construction and Interpretation of a Digital Thermal Inertia Image.
Photogrammetric Engineering and Remote Sensing. 1977; XL111(8): 938-999.
[4] SH Miller, Kennech Watson. Evaluation Algorithms for Geological Thermal-Inertia-Mapping.
Proceedings of the 11th International symposium on Remote Sensing of Environment. 1977; 11:
1147-1160
[5] Lu zifeng. Calibration and the measurement error analysis of infrared imaging system for temperature
measurement. Dissertation. Beijing: Chinese Academy of Sciences; 2009.
[6] Konrad JK, Buettner Clifford D Kern. The determination of infrared emissivities of terrestrial surfaces.
Journal of Geophysical Research. 1965; 70: 1329-1337.
[7] YU Da-bin. Effect of infrared emissivity of coatings on the camouflage effectiveness of targets.
Infrared and Laser Engineering. 2007; 36: 194-197.
[8] Li junshan. Infrared Image Processing, Analysis and Fusion. Beijing: Science Press. 2009: 11-13.
[9] Cheng HD, Li JG, Chen JR. Threshold selection based on fuzzy c-partition entropy approach.
Pattern Recognition. 1998; 31: 857-870.
[10] Jui-Cheng Yen, Fu-Juay Chang. A New Criterion for Automatic Multilevel Thresholding. IEEE
Transaction Image Processing.1995; 4: 370-378.
[11] Hanshan Li,Zhiyong Lei.Research on Infrared Special Facula View Measurement Method Based on
Image Processing Technology. TELKOMNIKA Indonesian Journal of Electrical Engineering. 2012;
10(6): 1422-1429.
[12] Li Zhengzhou etc.Gray-scale Edge Detection and Image Segmentation Algorithm Based on Mean
Shift. TELKOMNIKA Indonesian Journal of Electrical Engineering. 2013; 11(3): 1414-1421.
[13] Legault R, Suen CY. Optimal local weighted averaging methods in contour smoothing. IEEE
Transactions on Pattern Analysis and Machine Intelligence. 1997; 19: 801-817.
[14] Giordana N, Pieczynski W. Estimation of generalized multisensory hidden Markov chains and
unsupervised image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence.
1997; 19: 465-475.
[15] Zhang tan, Chen gang. Image processing based on partial differential equation. Beijing: Higher
education press. 2004.
[16] Tian Ying, Yuan Wei-qi. Application of the Genetic algorithm in image Processing. Journal of Image
and Graphics. 2007; 12: 389-396.
[17] Yansun Xu, Weave JB. Wavelet transform domain filters: a spatially selective noise filtration
technique. IEEE Transactions on Image Processing. 1994; 3: 747-758.
[18] Mallat S, Hwang WL. Singularity detection and processing with wavelets. IEEE Transactions on
Information Theory. 1992; 38: 617-643.
[19] Deng Shiwei, Yuan Baozong. Range Image Segmentation Based on Mathematical Morphology. Acta
Electronica Sinica. 1995; 23: 6-9.
[20] LUO Shu-qian, TANG Yu. A Bias Based Adaptive Fuzzy Segmentation Algorithm. Journal of Image
and Graphics. 1999; 7A: 111-114.
[21] XUE Jinghao, ZHANG Yujin, LIN Xinggang. New thresholding algorithm based on fuzzy divergence
for image segmentation. Beijing: Tsinghua Univ (Sci&Tech). 1999; 39: 47-50.
[22] Kuntimad G, Ranganath HS. Perfect image segmentation using pulse coupled neural networks.
IEEE Transactions on Neural Networks. 1999; 10: 591-598.

253
[23] Gonzalez HC. Digital Image Processing Using MATLAB. 3rd ed. Beijing: Publishing House of
electronics industry. 2005: 289-293.
[24] Qi wenjuan. Effect of emissivity on accuracy of infrared temperature measurement. Thesis.
Changchun university of science and technology. 2006.
[25] Electronic Publication: Wikipedia.
http://commons.wikimedia.org/wiki/File:Heatequation_exampleB.gif?uselang=zh.
[26] WANG Ran, PING Xi-jian. Steganalysis of LSB Matching Based on Local Smoothness of Histogram.
Journal of Applied Sciences in Shanghai. 2012; 30: 96-103.
[27] Jinwei Li, Guiping Liao. Crop Image Analysis System Program Ver2.0

Multi-Emissivity Setting in Thermal Imaging Based on Visible-Light Image Segmentation

More Related Content

What's hot (20)

Similar to Multi-Emissivity Setting in Thermal Imaging Based on Visible-Light Image Segmentation (20)

More from Nooria Sukmaningtyas (20)

Recently uploaded (20)

Multi-Emissivity Setting in Thermal Imaging Based on Visible-Light Image Segmentation