Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Procedia
Engineering
Procedia Engineering 00 (2011)29000000
Procedia Engineering (2012) 420 424
www.elsevier.com/locate/procedia

2012 International Workshop on Information and Electronics Engineering (IWIEE)

Infrared Small Targets Detection based on MRF Model

Xinde Suna*,Guizhen Fanga
a
Zhengzhou Institute of Aeronautical Industry Management,Zhengzhou 450015,China

Abstract

Aiming at the difficulty in detecting infrared dim small target under strong background clutter, we propose a novel
algorithm of infrared dim small target detection based on Markov random fields(MRF) model. We at first use
adaptive morphological filter to suppress the background clutter, then introduce new potential function and energy
function according to MRF theory and infrared small target image features, and build a target detection model to
confirm automatically the location and size of the target. Simulation results show that the algorithm is effective.

2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.

Kewwords:MRF; energy function; clutter suppression; infrared small targets

1. Introduction

At present, with the use of infrared stealth technology and the increased detection range, the radiation
intensity and radiation area of infrared target are drastically reduced. Thus, the target in the infrared
imaging plane showed only a few to a dozen pixels, Called small targets. It is an important basis for
infrared automatic targets recognition to detect quickly and accurately small targets drowned by the
background clutter and noise, whose performance will directly determine the system's detection
sensitivity and detection range.
The Markov random fields (MRF) model, as a statistical description method of image, does not require
the image is stationary or wide-sense stationary, is widely used in image reconstruction, image anti-drop
clarity, texture analysis and other fields[1].However, MRF applied research in the field of infrared dim
small target detection is still relatively small.
In view of this, an infrared image small target detection method based on the MRF model is presented

* Xinde Sun. Tel.: 0371-56865312;

E-mail address:sunxinde@zzia.edu.cn

1877-7058 2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
doi:10.1016/j.proeng.2011.12.734
2 Xinde Sun and Guizhen Fang/
Fang / Procedia
Procedia Engineering
Engineering 00
29 (2011)
(2012) 000000
420 424 421

in this paper. Experimental results show that this method could detect the infrared small target drowned
by the background clutter and noise.

2. The principle of small target detection based on MRF mode

Although the small targets in infrared low signal to noise ratio image are often overwhelmed by the
background clutter, but there are still differences between small targets and background on the gray, and
after clutter suppression and target enhancement treatment, the differences are further amplified. In the
image gray space, a small target corresponds to a number of neighbors singular point; therefore, there is
a special relationship between the target pixel and its neighborhood pixels. We use the MRF theory of
Clique and Energy Function to describe this relationship in this paper, and separate targets from
background by quantifying and calculating this relationship, thus achieve the target detection.

2.1. MRF neighborhood systems and clique concept [2,3]

Fig.1. Clique structure types of the second-order neighborhood system

In the MRF theory, For a size of M N image S = {s (i, j );1 i M ,1 j N } , if mij and pixel s(i,j)
meet relations: s (i, j ) mij ,and if s (k , l ) mij ,then s (i, j ) m kl , then we can call mij a neighborhood of s
{
(i, j).Hence neighborhood system can be expressed as: M = mij (i, j ) M N , mij M N . }
Clique and the neighborhood system M are related, it is a subset of S, shows the interaction between
image pixels and texture composition. A second-order neighborhood system defined by Euclidean
distance may have 10 types of clique structure, which are shown in Fig. 1.We denote c as a clique, and
denote C as all cliques belonging to a second-order neighborhood system.

2.2. Definition of potential function and energy function for small target detection

Potential function and energy function are two important concepts based on clique. In different
application environments, potential function and energy function have usually different definition form,
and that is the key to successful application [2,3].

2.2.1. Definition of potential function

Potential function is defined as statistics of the gray difference between a pixel and pixels of its clique,
which reflects the gray change in cliques of the pixel. According to second-order neighborhood system
clique structure, we divide cliques set C into three subsets C1, C2, C3, and define three corresponding
potential function. C1 represents a subset of vertical or horizontal two-point structure cliques; C2
422 XindeXinde SunGuizhen
Sun and and Guizhen
Fang/Fang / Procedia
Procedia Engineering
Engineering 29 (2012)
00 (2011) 420 424
000000 3

represents a subset of diagonal two-point structure cliques; C3 represents a subset of three-point structure
cliques; single-point and four-point structure cliques would not be taken into account.
For C1, by comparing respectively the gray difference between center pixel and four neighboring
pixels with the threshold T1, we can get four logic values:
1, abs( s(i, j ) s(i 1, j )) T 1 1, abs( s(i, j ) s(i, j 1)) T 1
delta1 (i, j ) = delta 2 (i, j ) =
0, abs( s(i, j ) s(i 1, j )) < T 1 0, abs( s(i, j ) s(i, j 1)) < T 1
1, abs( s (i, j ) s (i + 1, j )) T 1 1, abs( s(i, j ) s(i, j + 1)) T1
delta3 (i, j ) = delta 4 (i, j ) = (1)
0, abs ( s (i , j ) s (i + 1 , j )) < T 1 0, abs ( s(i, j ) s(i, j + 1)) < T 1
Then the potential function is defined as :VC1=1,if Sum(deltak(i,j)T2,k=1,2,3,4;VC1=0, else. (2)
For C2, comparing respectively the gray difference between center pixel and four neighboring pixels
with the threshold T2, we can get four logic values:
1, abs(s(i, j) s(i 1, j 1)) T 3 1, abs(s(i, j) s(i + 1, j 1)) T 3
delta1(i, j) = delta2 (i, j) =
0 , abs(s (i, j ) s (i 1 , j 1 )) < T 3 0, abs(s(i, j) s(i + 1, j 1)) < T 3
1, abs(s(i, j) s(i + 1, j + 1)) T 3 1, abs(s(i, j ) s(i 1, j + 1)) T 3
delta3 (i, j) = delta4 (i, j ) = (3)
0, abs(s(i, j) s(i + 1, j + 1)) < T 3 0, abs(s(i, j ) s(i 1, j + 1)) < T 3
Then the potential function is defined as: VC2=1,if Sum(deltak(i,j)T4,k=1,2,3,4;VC2=0, else. (4)
For C3, comparing respectively the gray difference between center pixel and four neighboring pixels
with the threshold T5, we can get four logic values:
1, abs(s(i, j) max(s(i 1, j), s(i 1, j 1), s(i, j 1))) T 5
delta1 (i, j) =
0, abs(s(i, j) max(s(i 1, j), s(i 1, j 1), s(i, j 1))) < T 5
1, abs(s(i, j) max(s(i, j 1), s(i +1, j 1), s(i +1, j))) T 5
delta2 (i, j) =
0, abs(s(i, j) max(s(i, j 1), s(i +1, j 1), s(i +1, j))) < T 5
1, abs(s(i, j) max(s(i +1, j), s(i +1, j +1), s(i, j +1))) T 5
delta3 (i, j) =
0, abs(s(i, j) max(s(i +1, j), s(i +1, j +1), s(i, j +1))) < T 5
1, abs(s(i, j) max(s(i, j +1), s(i 1, j +1), s(i 1, j))) T 5
delta4 (i, j) = (5)
0, abs(s(i, j) max(s(i, j +1), s(i 1, j +1), s(i 1, j))) < T 5
Then the potential function is defined as: VC3=1,if Sum(deltak(i,j)T6,k=1,2,3,4;VC3=0, else. (6)

2.2.2. Definition of energy function

Energy function is defined as the sum of the potential functions, that is:
3
U C (i, j ) = V (i, j ) = V
cC
c
k =1
Ck (7)

Energy function value reflects the pixel gray-scale roughness of the neighborhood, which is the basis
for target detection.

2.3. Target detection idea

Infrared image small target displays a bright plaque composed of several pixels. Target plaque
boundary could be composed some of gray singular points, the roughness of neighborhood of these
singular points must be very high. However, the roughness of neighborhood of those pixels in
background area or plaque central area will be relatively low. Therefore, according to definition of energy
4 Xinde Sun and Guizhen Fang/
Fang / Procedia
Procedia Engineering
Engineering 00
29 (2011)
(2012) 000000
420 424 423

function given above, the energy function value of pixels in target plaque boundary must be bigger. In
other words, target plaque boundary must be included in high-energy pixels.
We set the appropriate energy threshold, and extract all high-energy pixels from an image, then
classify the pixels by Euclidean distance. If the Euclidean distance between two pixels is less than the
threshold, we can think they belong to the same pixel group. Obviously there may be multiple pixel
groups in an image. If a pixel group contains more pixels, then we can consider the pixel group
corresponding to a target plaque.
This target detection method can remove some salt and pepper noise, which can make the detecting
results more reliable.

3. Infrared small targets detection algorithm based on MRF model

We can get the target detection algorithm steps based on the above analysis:
Suppress the background clutter and enhance the target by adaptive morphological filtering from [4].
Use (7) to compute the energy function values corresponding to each pixel beyond the image borders.
Extract the high-energy pixels by setting the appropriate energy threshold.
Classify the picked pixels by Euclidean distance, identify the target plaques in the image, and deter -
mine the location and number of targets.

4. Experimental results and analysis

On MATLAB R2006 platform, we have tested the algorithm based on real images in different
environment.
In our experiments, image 1 is the infrared sky background image, size of 97113; the target is a plane
through the clouds, size about 35. Image 2 is the infrared plane landing image, size of 148168, where
the target size of 1116; background is composed of sky, earth, trees and infrared noise. Image 3 is water
target infrared image whose background is sea and sky, size of 110144, where the target size of 49.
The parameters value associated with the experiment were: T1=0.5s(i,j), T2=2, T3=0.5s(i,j), T4=2,
T5=+10-4, T6=1, the threshold of energy function UC is 3. T1 and T3 are the relative value. The other
parameters are constant values. We use the Euclidean distance threshold of 15 to classify the high-energy
pixels. When a pixel group contains 5 or more high energy pixels, we believe that a target plaque is
detected.

(a) (b) (c)

Fig.2 (a) comparison of image 1 before and after processing;(b) comparison of image 2 before and after processing;(c) comparison
of image 3 before and after processing
424 XindeXinde SunGuizhen
Sun and and Guizhen
Fang/Fang / Procedia
Procedia Engineering
Engineering 29 (2012)
00 (2011) 420 424
000000 5

Fig.2 shows that the adaptive morphological filter has satisfactory results of background clutter
suppression, target has been significantly enhanced which would ensure the reliability of target detection.
Signal to noise ratio (SNR) in the table 1 is defined by local least variance method [5]. Experimental
results show that this algorithm can effectively detect the target even in the complex environment of low
SNR, and can determine accurately the target location and size.

Table 1. Target detection experiment results

Original information Test results

SNR Location Size Location Size
image 1 12dB (60,53) 35 (61,52) 36
image 2 10dB (86,48) 1116 (85,50) 1118
image 3 5dB (51,70) 49 (53,70) 49

5. Conclusion

In this paper, we use the MRF theory to describe the relationships between pixels in the image, and
define a new energy function computational model according to the characteristics of infrared small target
image, then establish pixels classification and target plaques recognition model. The algorithm enhances
the effect of complex background clutter suppression because of use of adaptive morphological filter
based on image local features, enhances the model and the actual degree of compliance because of
establishing a target detection model based on MRF model, so ensure the sensitivity and reliability of
small target detection. Simulation experiment depending on real images and empirical parameters shows
that the proposed algorithm is effective.

Acknowledgements

This work was supported by National Natural Science Foundation of China (No.41001235) and
Aeronautical Science Foundation of China (No. 20095155008).

References

[1] R.Pan,S.J.Reeves. Efficient huber-Markov edge-preserving image restoration[J].IEEE Transactions on Image Processing,
2006,15(12): 3728-3735
[2] Y. Huang,K. Fu,Y. R. Wu. Image segmentation method using K-Means based on Markov random field[J]. Acta Electronica
Sinica, 2009,37(12):2700-2704. (in Chinese)
[3] D. B. Wang,S. Q. Liu,X. M. Kou,et al. Infrared background clutter suppression algorithm of adaptive regularization based on
MRF[J]. J. infrared millim. Waves,2009,28(6):440-444. (in Chinese)
[4] X. D. Sun,S. K. Bo,L. L. Li. A study of infrared image clutter suppression based on background estimation[J].Laser &
Infrared,2011,41(5):586-590. (in Chinese)
[5] X. W. Wang,C. X. Wang,Y. Y. Zhang. Research on SNR of point target image[J]. Electronics optics
control,2010,17(1):18-21. (in Chinese)