Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h, Vo lume 3, Issue 3, Marc h-2012 1

ISS N 2229-5518

The Segmentation of FMI Image Layers Based on FCM Clustering and Otsu thresholding

J. Gholampour, A.A. Pouyan

Abs tractA key aspect in extracting quantitative inf ormation from FMI logs is to segment the FMI image to get image of layers. In this paper, an automatic method based on FCM clustering and Otsu thresholding is introduced in order to extract quantitative inf ormation from FMI images. All pixels are clustered using FCM clustering algorithm at the f irst step. The second step uses KNN f or other clus tering. Then, uncovered columns of FMI image and image inequality are removed. Finally, the Otsu thresholding method is investigated f or improving pixel-clustering step. Filed data processing examples show that sub image of layers can be accurately seprated f rom original FMI images.

Inde x TermsFMI, Seg mentation, Otsu, FCM, Thresholding, KNN, Clustering

—————————— ——————————

1 INTRODUCTION

He difficulty in carbonate reservoir evaluation come from the complexity of reservoir storage spaces (such as pores, vugs, fractures and their combination) and thir hetero- geneity. Because of fractures, the validity of the reservoir is not easily identified by conventional log. The imaging log provides a new means for such a complex reservoir evalua- tion. FMI (Full Micro-resistivity Imager) provides borehole wall resistivity imaging. The FMI tool is made of eight elec- trode pads with 24 button electrodes on each pad. The FMI tool sample interval in depth and well circumference is 0.1 inch. The tool covers approximately 80% of the borehole wall in an 8.5 inch hole. The FMI image clearly shows the geologi-
cal phenomena around the borehole wall [1].
In order to extract qualitative information from FMI, a basic
step is used to segment the FMI image to get sub-image of layers. Then, the segmented images are analyzed and processed to extract relevant information. The segmentation result directly affects the accuracy of parameter calculation [2].
A number of image segmental algorithms have been devel-
oped in the literature. They can be rouly grouped in two cate-
gories: area description-based algoriths and edge detection- based methods [3].
In this paper, we present a segmentation method based on
Fuzzy C-mean algorithm and Otsu thresholding method. The

2 P ROPOS ED M ETHOD

The complete block diagram representation of the proposed method is shown in Figure 1. After inputting a FMI image, we select intensity values of red, green, blue and equivalent gray level of each pixel as features. Then we normalize these fea- tures for each row of image. At section 2.1 we illustrate the normalization method. After this step, we cluster pixel of FMI image using fuzzy c-mean (FCM). Feature selection, feature normalization and FCM have been performed to last row of FMI image.
Then k–nearest neighbor algorithm (KNN) has been called
on each row of image. KNN has been explained in section 2.3.
Then uncovered columns have been removed (section 2.4). We illustrate removing image inequality method in section 2.5. In section 2.6 we explain Otsu thresholding method.

2.1 Normalization

As leave-one-out scheme is used, for each feature in the train- ing samples the feature normalization is adapted based on the following scheme:

f

first step classify pixel of each row of FMI image using fuzzy C-mean algorithm. At the second step, we use k-nearst neigh- bor for classifying. Then we remove uncovered columns and

f i i

i

(1)
remove image inequalities. At the end we use Otsu threshold-
ing method.

———— ——— ——— ——— ———

Ph.D., Asst. Professor Computer Engineering, School of Computer Eng i- neering, Shahrood University of Technology, Shahrood, Iran. E -mail: apouyan@ieee.org

Graduate Student Computer Engineering, School of Computer Engineer- ing, Shahrood University of Technology, Shahrood, Iran. E-mail: jgholam- pour@gmail.com

Where fi, µi and σi are respectively i-th feature, mean and
standard deviation of i-th feature [4].

2.2 Fuzzy C-Mean s Clu ster

Feature selection and feature normalization run on each row of FMI image, then fuzzy c-means (FCM) executed.This process runs until arrive to last row of FMI image (Fig 1).
FCM is a method of clustering which allows one piece of
data to belong to two or more clusters. This method is fre- quently used in pattern recognition [5]. It is based on minimi- zation of the objective function:

IJSER © 2012

http :// www.ijser.org

Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h Vo lume 3, Issue 3 , Marc h-2012 2

ISSN 2229-5518

n c each member of the training set is to the test class that is being

J q (U ,V )



k 1 i 1

(uik ) d

( xk , vi )

(2)

examined. Euclidean Distance measuring:

N

2 2 (8)

where X ={x1, x2,…, xn}, n is the number of data items, c is the number of clusters with 2 ≤ c ≤ n, uik is the degree of mem- bership of xk in the i-th cluster, q is a weighting exponent on

d E (x , y ) 

i 1

x i y i

each fuzzy membership, vi is the prototype of the center of cluster i, d2 (xk, vi) is a distance measure between xk object and

cluster centre vi. A solution of the object function can be ob- tained via an iterative process, which is carried as follows:
1) Set value for c, q and ε.
2) Initialize the fuzzy partition matrix U.
3) Set the loop counter b = 0.
4) Calculate the c cluster centers {vi(b)} with U(b):
From this k-NN category, class label of the test pixel image is

determined by applying majority voting [8].

Start

n

(u ik ) x k

Last row of FMI image?

(b ) k 1

i n

(3)

Yes

k 1

(b ) q ik

NO

5) Ca lculate the me mbers hi U(b+1). For k = 1 to n,calcu late
the following

Feature selection

For the k-th column of the matrix, compute new membership values :
If Ik = ᵠ , then

Feature

Normalization

Fuzzy C-Mean Clustering

(b 1) 1

ik c d

(6)

( ik )2 (q 1)

j 1

d jk

KNN

(Number of rows of FMI Image times)

Else uik(b-1) = 0 for all I Ik and

u (b 1)  1

(7)

Remove uncovered columns from

ik

i I k

next k

FMI Image

6) If U (b ) U (b 1)

and go to s tep 4 [6].

, s top; otherwis e, s et b=b+1

Removing image inequality

2.3 K-Nearest Neighbor

K-nearest neighbor (k-NN) is a supervised learning algo- rithm by classifying the new instances query based on majori- ty of k-nearest neighbor category. Minimum distance between query instance and the training samples is calculated to d e- termine the k-NN category. The k-NN prediction of the query instance is determined based on majority voting of the nearest neighbor category. Since query instance (test image pixel) will compare against all cluster [7].
In this works, for each test pixel image (to be predicted),
minimum distance from the test pixel image to the training set is calculated to locate the k-NN category of the training data set. Euclidean Distance measure is used to calculate how close

Otsu thresholding

Finish

Fig. 1. Block diagram of proposed system

2.4 Removing Un covered Column s

At this step, we assume the columns that all pixels categorized to same class, are uncovered columns of FMI image (20% of FMI image). Therefore these columns remove.

IJSER © 2012

http :// www.ijser.org

Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h Vo lume 3, Issue 3 , Marc h-2012 3

ISSN 2229-5518

2.5 Removing Image Inequality

In proposed system, a nonlinear filter is used that remove non-edge surface roughness image and keep edges FMI images. First, for each pixel we calculated eight derivatives in all direction. Then the combination of the eight derivatives added to pixels value. Figure 2 shows the pixel (n, m) and its eight neighbors. Equation (9) computed the derivative of pixel
neighbors. In this equation, I i represents the gray level of i-th neighbors of pixel (n, m). After calculating derivatives in all direction, these values multiplied by the coefficient δ and added to pixel value (equation (10)). As shown in equation (9), this filter has three parameters (K, α and δ). Their value in this study is considered 0.02, 2.5 and 1.25 respectively.
The means of class C1 and C2 are


The total mean of gray levels is denoted by uT

Figure 2: pixel (n, m) and 3 ×3 neighborhood

uT w 1u1 w 2u 2

(16)

i I i I (n, m)

(9)

The class variances are

I (n, m)  I (n, m)  { 1



   }

1  1

 

1  8

(10)

2 2 i 2

K K

The within -class variance is

M

As noted above, this filter removes non-edge surface rough-
ness and keeps edges in FMI images. Based on above relation- ships, if derived value in the edge pixel be larger than parame-

2 2

k k k 1

(19)

ter K then the absolute operand is greater than 1. After ^, ob-
The between-class variance is
tained a large value and changed a little on the intensity of edge pixel. Therefore the filter has no effect on the edge pixels.

2 w

1 u1

uT

2 w

2 u 2

uT

(20)

Furthermore, for non-edge pixels, the absolute value is smaller
than 1 and can skip it. Thus equation (10) in limit case is Lap-
The total variance of gray levels is

2 2 2

(21)

lace for non-edge pixels and acts as a smoothing filter. In addi- tion, we can use this nonlinear filter much time [8]. Figure 7 shows multiple repeated of this filter and in each stage, the image obtained is displayed [9].

2.6 Otsu Method

In this section, we use Otsu method [10] as follows. Assuming

T w B

Otsu method chooses the optimal threshold t by maximizing the between-class variance, which is equivalent to minimizing the within-class variance, since the total variance (the sum of the within-class variance and the between-class variance) is constant for different partitions [11].
an image is represented in L gray levels [0, 1, …L-1]. The

t  arg{ max {2 (t )}}  arg{ min {2 (t )}}

(22)

number of pixels at level i is denoted by ni, and the total num-

B

0t L 1

w

0t L 1

ber of pixels is denoted by N = n1 + n2 +...+ nL . The probability of gray level i was denoted by

L 1

Otsu method can be extended to multilevel thresholding me- thod. Assuming that there are M-1 thresholds [t1,t2,…,tM-1] that divide the pixels in the image to M classes{C1, C2,…, CM}

pi ni / N , pi  0, 0

pi  1

(11)

t , t ,, t

 arg{ max {2 (t ,t

,...,t

)}}

1 2 M 1

2

0t L 1

B 1 2

M 1

(23)

In bi-level thresholding method, pixels divided into two classes C1 with gray levels [0, 1, …, t] and C2 with gray levels [t+1, …, L-1] by the threshold t . The gray level p robability dis- tributions for two clas s es are

 arg{ min {W (t1 ,t 2 ,...,t M 1 )}}

0t L 1

Where

IJSER © 2012

http :// www.ijser.org

Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h Vo lume 3, Issue 3 , Marc h-2012 4

ISSN 2229-5518

KNN Classifier, International Conference on Intelligent and Ad- vanced Systems(ICIAS), IEEE, 2010

[9] R.C.Gonzalez and R.E .Woods, ―Digital image processing‖, 2nd Ed i- tion, Prentice-Hall 2002. N.

[10] Otsu, ―A threshold selection method from gray-level histogram ‖,

IEEE Transactions on System Man Cybernetics, Vol. SMC-9, No. 1:

62-66, 1979.

[11] S.S. Reddi, S.F. Rudin, and H.R. Keshavan, ―An optimal multiple

threshold scheme for image segmentation,‖ IEEE Trans. System Man

Cybernet. 14(4): 661 –665, 1984.

3 EXPERIM ENTS AND R ESULTS

Original FMI image is represented in Fig. 3. The output of all steps of proposed system has been shown in Fig. 4 to Fig. 8. Fig. 4 is pixels clustering after FCM algorithm. The result of KNN algorithm is shown in Fig. 5. Removeing of uncovered columns of FMI image is determind in Fig. 6. Fig. 7 shows Output image after Removing Image Inequality. Finally, Fig. 8 shows the resulting segmented image after Otsu thresholding method.

4 CONCLUSION

We have proposed a segmentation method based on FCM clustering and Otsu thresholding. The field data processing examples show that sub image of layers can be accurately s e- prated from original FMI images.

REFERENCES

[1] S.M. Luthi, ―Geological well logs: their use in reservoir modeling‖,

Springer, 1990.

[2] S.K. Pal, J.F. Peters, ―Rough Fuzzy Image Analysis: Foundations and

Methodologies‖, CRC Press.

[3] L.R. Lin, W.Y. Qi, L. J. Hua, and M. Yong, ―The Segmentation of FMI Image Based on 2-D Dyadic Wavelet Transform‖, Geophysics, Vol. 2, No. 2, June 2005.

[4] A.A. pouyan, H. Hassanpour, and H. Dehghan, ―SVM-based Diag- nosis of the Alzheimer’s Disease using 18F-FDG PET with Fisher Dis- criminant Rate‖, The 18th Iranian Conference of Biomedical Engi- neering (ICBME2011).

[5] J. C. Bezdek, ―Pattern Recognition with Fuzzy Objective Function

Algoritms‖, Plenum Press, New York, 1981.

[6] Z.K. Huang, P.W. Li, S.Q. Wang, and L.Y. Hou, ―Using FCM for Col- or Texture Segmentation Based Multirscale Image Fusion‖, Interna- tional Conference on e-Education, e-Business, e-Management and e- Learning, IEEE 2010.

[7] P. Pallabi, T. Bhavani, ―Face Recognition Using Multiple Classifiers‖,

18th IEEE International Conference on Tools with Artificial Intelli-

gence, pp. 179-186, 2006.

[8] M.N. Mansor, S. Yaacob, R. Nagarajan, L.S. Che,M.Hariharan, and

M. Ezanuddin , Detection of Facial Changes for ICU Patients Using

IJSER © 2012

http :// www.ijser.org

Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h Vo lume 3, Issue 3 , Marc h-2012 5

ISSN 2229-5518

Fig. 3. Original FMI image.

Fig. 4. Pixels af ter FCM clustering algorithm.

Fig. 5. Pixels af ter KNN algorith m.

Fig. 6. The uncovered columns of FMI image are determind.

Fig. 7. Output image af ter Removing Image Inequality.

Fig. 8. Final output af ter Otsu thresholding method.

IJSER © 2012

http :// www.ijser.org