An Approach of Segmentation Technique of SAR Images using Adaptive Thresholding Technique

An Approach of Segmentation Technique of SAR Images using Adaptive Thresholding Technique

Debabrata Samanta1, Goutam Sanyal2

Department of CSE

National Institute of Technology, Durgapur, West Bengal, India – 713209

Abstract

Segmentation is an essential matter in synthetic aperture radar (SAR) image analysis. When contrast results against the true forest, ocean, and hills etc SAR images can afford more information, and obtain more acceptable segmentation than only single polarization SAR image. In this paper a novel methodology has been carried out to segment a SAR image based on the Adaptive Thresholding Technique. This technique composes of three main processes: firstly, selecting training samples i.e. mean value for every region in the SAR image. Secondly, training these samples using 5X5 window, and obtain variance of every region. Finally, the segmentation of SAR image with respect to group threshold value i.e. generated automatically.

Keywords

SAR image , Gray Image, Adaptive Thresholding.

1. Introduction

   A SAR image does not depend on specification or light conditions and images can be taken at any part of the day and in any specification. During the processing of SAR image, a correct segmentation lays the foundation for the following analysis, such as feature extracting, classifying and extracts the information of geographical areas. A large number of algorithms have been proposed to contend with the segmentation problem in SAR images.

2. Related Work

   The problem related to the segmentation of SAR images is depending on the intensity correlation. There are many scholastic researches, who have paved their ways for SAR images segmentation.

   Yongmin Shuai , Hong Sun, Ge Xu proposed SAR image segmentation Based on Level Set i.e. is the computational expense inherent to solving the proposed nonlinear parabolic partial differential equation [1]. Xiaolin Tian, Shuiping Gou and Licheng Jiao developed SAR image segmentation technique based on spatially adaptive weighted Possibilistic c-means clustering [2]. Ming Li Yan Wu Shunjun Wu proposed Segmentation technique based on Mixture Context and Wavelet Hidden-Class-Label Markov Random Field [3]. Sun Li, Zhang Yanning, Ma Miao, Tian Guangjian proposed infinite mixture model can simulate the intrinsic property of SAR image and the segmentation method can determine the cluster number automatically [4]. Haiyan Li, Zhengyao Bai describes a new PCNN-Based method for Segmentation of SAR Images [5]. Ting Liu, Xian- Bin Wen, Jin-Juan Quan and Xue-Quan Xu integrate the multi-scale technology, mixed-model information and support vector machines (SVM) for SAR image segmentation [6]. Xueying Yan, Licheng Jiao, Shuwen Xu proposed segmentation method based on gabor filters of Adaptive window of SAR image [7]. Fucheng Li, Jin Zhang, Lili Zhou, Jianping Du, Bin Yan used a new segmentation method of SAR image based on multi-polarization information fusion [8]. Han Ping, Zhang Rui, Su Zhi-gang, Wu Ren-biao proposed segmentation method based on Support Vector Machine [9].

   In this paper a novel methodology has been carried out to segment a SAR image based on the Adaptive Thresholding Technique. This technique composes of three main processes: firstly, selecting training samples i.e. mean value for every region in the SAR image. Secondly, training these samples using 5X5 window, and obtain variance of every region. Finally, the segmentation of SAR image with respect to group threshold value i.e. generated automatically.

3. Proposed Methodology

Since the unexpected changes in intensity pixel value indicates class, its detection in binary or segmented

SAR=

2

Q µ

r ,c SAR

w

5

2

image is quite straightforward. However, the process of class is less complex in the case of SAR images. As an initial approximation, the number of segments is pre-defined as W, and as many different, equally spaced mean values are obtained based on the maximum and minimum gray level values in the image, as given in the following equation USAR=(r*c)+minSAR, ————-(1)

Where, µ, is the mean of the ith class, QSAR=( max- min)/(W-1) —————(2)

max is the maximum gray level intensity value, and

Where 2SAR the variance or square of the standard

deviation, and w is the number of samples in the region. An estimate of the difference in the variance values of the region between the predicator and post- disaster images is an indication of the level of change. In our methodology first we have convolute the image with a five order mask according to the X and Y direction respectively to obtain the gradient of each points. Then the ultimate gradient is obtained by the following formula: f(r,c) = [ Gr Gc ] = [ f/x

f/y ] ——–(6)

The weight of the vector is: f(r,cy) = mag(f(r,c))

min is the minimum gray level intensity value. Each

=(G2 +G2

) ———(7)

r c

of the pixel values in the image is now assigned to

the bin of the corresponding class based on their distance from the mean value of each class. As a measure to identify the distance, the linear distance is calculated as:

Gradients of points give the maximum location at the center of the original signal thus removing the noises from SAR images. Then the gradient of the pixels is calculated as:

distSAR =Yr,c – QSAR —————-(3)

Where Yr,c, is the pixel intensity at the location (r,c).

Vr,c 1

15

5/2 5/2

QSAR (r,c)* FSAR (r h,c k) (8)

p/2 k 5/2

3. A. Mean: The statistical mean gives us the average value of all the pixels for a particular feature. Mathematically, the statistical mean of a set of variables is expressed as:

µSAR= Qr,cP Qr,c 4

The summation is carried out over all the gray level values in the set. It is important to notice that, even though the mean gives a good estimate of the characteristics of a feature, it might not give a good representation of the damage or change in the region.

3. B. Variance: Variance is the second moment of a matrix and is a measure of gray-level contrast that can be used to establish descriptors of relative smoothness. Mathematically, it is defined as the mean of squares of differences between respective samples (gray-level intensities) and their mean and is expressed as:

Where * indicates a discrete convolution, QSAR is the 5X5 mask, and FSR(r,c) is a hXk SAR image.

3. C. Adaptive Thresholding

Image classification based on thresholding is a greedy technique because it classifies pixels into two categories.

Category1: Pixels whose values fall below the threshold.

category2: Pixels whose values are equal or exceed the threshold.

If T is a threshold value, then any pixel (r, c) for which f(r,c) T is called a class point. For thresholding we compute adaptive threshold of local intensity variations as:

1. First the overall mean value of the SAR image is calculated. So the pixels having lower class strength

than this mean value are already discarded.

5. Experiment and Analysis

1 1 1

9

T1 V (r h, c K ) (9)

k 1 p

Threshold1 = mean (G).

2. Then a 3 X 3 window is splits over the SAR image where the mean and variance of the SAR image within this window are calculated.

In this thesis work, we have considered synthetic aperture radar images. The SAR images are segmented by using Adaptive Thresholdig Technique. The figure (Fig 1 to Fig 3) shows the original SAR images and segmented SAR images based on Adaptive Thresholding Technique.

SAR

1 G(r h, c k) T1 (10)

1

1

9

k 1 p

Then taking the sum of this mean value and standard deviation and this is considered as the threshold value of that pixel. Now if the gradient of this image exceeds this threshold then the pixel is treated as a class. At last this mean and standard deviation is summing up to generate the ultimate threshold for the candidate pixel,

TSAR=T1+SAR. —————(11)

In this way threshold value is generated dynamically and region wise for each pixels so that the possibility of data loss or noise is quite reduced. If the pixel value is greater than or equal to this adaptive threshold then only the pixel is treated as class otherwise it is discarded.

1. Proposed Algorithm

   • Input: SAR Images of variable size.

   • Output: Classification of SAR image.

1. Start.

2. Taken a SAR images.

3. Consider a 5X5 window.

4. calculate the mean, variance of that SAR Images.

5. Segmentation is obtained using adaptive threshold change.

6. Stop.

Figure 1. Input SAR image and Segmented SAR image

Figure 2. Input SAR image and Segmented SAR image

Figure 3. Input SAR image and Segmented SAR image

1. Conclusion

   In this paper, we proposed a novel Adaptive Thresholding Technique based image segmentation technique for SAR image. This technique based on considering a window and calculates the mean, variance of that SAR Images then store the color feature for Adaptive Thresholding. Next, Segmentation is obtained using Adaptive Thresholding value using region merging. For this

   reason, our algorithm did not make inaccuracy; that is, a segmented SAR image very different to get the originality of the SAR images. This may be extended to the color image segmentation. The results from this preliminary study indicated that the proposed strategy was effective.

2. References

1. Yongmin Shuai, Hong Sun, Ge Xu , A Fast Segmentation Scheme Based on Level Set for SAR Images, pp. 503-506,1-4244-1188-2/07/$25.00 ©2007 IEEE.

2. Xiaolin Tian, Shuiping Gou and Licheng Jiao , SAR image segmentation based on spatially adaptive weighted Possibilistic c-means clustering, Proceedings of 2007 International Symposium on Intelligent Signal Processing and Communication Systems ,pp. 304-307,Nov.28-Dec.1, 2007 Xiamen, China.

3. Ming Li Yan Wu Shunjun Wu , SAR Image Segmentation Based on Mixture Context and Wavelet Hidden-Class-Label Markov Random Field , Fourth International Conference on Fuzzy Systems and Knowledge Discovery (FSKD 2007).

4. Sun Li, Zhang Yanning, Ma Miao, Tian Guangjian , SAR Image Segmentation Method Using DP Mixture Models, pp. 598-601,978-0-7695-3498-5/08 $25.00 © 2008 IEEE.

5. Haiyan Li, Zhengyao Bai, A New PCNN-Based Method for Segmentation of SAR Images , 10th Intl. Conf. on Control, Automation, Robotics and Vision, Hanoi, Vietnam, pp. 1635- 1639,1720 December 2008.

6. Ting Liu, Xian-Bin Wen, Jin-Juan Quan and Xue-Quan Xu, Multiscale SAR Image Segmentation Using Support Vector Machines, 2008 Congress on Image and Signal Processing, pp. 706- 709,978-0-7695-3119-9/08 $25.00 © 2008 IEEE.

7. Xueying Yan, Licheng Jiao, Shuwen Xu , SAR Image Segmentation Based on Gabor Filters of Adaptive window In Over complete Brushlet Domain, pp. 660-663, 978-1- 4244-2732-1/09/$25.00 ©2009 IEEE.

8. Fucheng Li, Jin Zhang, Lili Zhou, Jianping Du, Bin Yan, A New Segmentation method of SAR Image Based on Multi-polarization information Fusion, 2009 Fifth International Conference on Natural Computation , pp. 99- 103, 978-0-7695-3736-8/09 $25.00 © 2009 IEEE.

9. Han Ping, Zhang Rui, Su Zhi-gang, Wu Ren-biao, An Iterative Segmentation Algorithm of SAR Image

   Based On Support Vector Machine, pp. 676-679, 978-1- 4244-2732-1/09/$25.00 ©2009 IEEE.

10. Debabrata Samanta, and Goutam Sanyal, A Novel Statistical Approach For Segmentation Of Images, Journal of Global Research in Computer Science (JGRCS), pp. 9- 13, Volume 2, No. 10, October 2011, ISSN: 2229-371X.

11. S.R. Parker, A.H. Kayran . "Lattice Parameter Autoregressive Modeling of Two- Dimensional FieldsPart I: The Quarter-Plane Case", IEEE Trans. on Acoust. Speech, Sig. Process., Vol 32, No.4, pp. 872-874, 1984.

12. N.H. Kaplan, Erer, S. Kent, "Fusion of Remote Sensing Images via Lattice Filters", Proceedings of 3rd Int. Conf. on Recent Advances in Space Technologies RAST 2007, 285-288, 14-16 June 2007, Istanbul, TURKEY.

Debabrata Samanta is a member of the IAENG, Board member of the Seventh Sense Research Group Journals (SSRGJ), member of Editorial Board of IJSCE. He obtained his B.Sc. (Physics Honors) in the year 2007, from the

Vivekananda Collage, Takurpukur, under Calcutta University; Kolkata, India .He obtained his MCA in the year 2010, from the Academy Of Technology, under WBUT. He is working his PhD in Computer Science and Engg. In the year 2010 from National Institute of Technology, Durgapur, India in the area of Image Processing .His areas of interest are Artificial Intelligence, Natural Language Processing and Image Processing. He has guided 10 PG and 20 UG thesis. He has published 37 papers in International Journals / Conferences.

Gautam Sanyal is a member of the IEEE. He has received his B.E and M.Tech degree from National Institute of Technology (NIT), Durgapur, India. He has received Ph.D. (Engg.) from

Jadavpur University, Kolkata, India, in the area of Robot Vision. He has published nearly 110 papers in International and National Journals / Conferences. 3 Ph.Ds (Engg) have already been awarded under his guidance. At present he is guiding eight Ph.Ds scholars in the field of steganography, Cellular Network, High Performance Computing and Computer Vision. He is presently working as a Professor in the department of Computer Science and Engineering and also holding the post of Dean (Students Welfare) at National Institute of Technology, Durgapur, India.