Professional Documents
Culture Documents
1
Mohammad Reza Yazdchi, 2Arash Golibagh Mahyari, 3Ali Nazeri
1
Department of Biomedical Engineering, Faculty of Engineering, Isfahan University, Iran
2
Department of Electronic and Electrical Engineering, Faculty of Engineering, Shiraz
University, Iran
3
Mobarakeh Steel Company, Isfahan, Iran
1
yazdchi@eng.ui.ac.ir, 2mahyari.arash@gmail.com, 3nza@mobarakeh-steel.ir
4. Defect Detection
To find the position of defect in an image, the local
entropy of the pixels of an image is calculated. The
Figure 1. A sample image of Sticking defect before local entropy of a pixel is defined as the entropy of the
enhancement pixels in a 9×9 matrix around it. The local entropy
matrix for the sample image in Figure 4 is calculated
and shown in Figure 5. Considering an appropriate
threshold, the local entropy matrix is converted to a
binary image. This threshold is computed adaptively
using Otsu’s method [3]. The binary image is
presented in Figure 6. In order to eliminate nonrelated
regions, we use morphological approaches [4,5].
Subsequently the binary image is eroded by a disk
structure element which has a radius of 5 pixels. So,
areas of all white regions decrease uniformly. In the
next stage all regions with an area below 2000 pixels
are deleted. We then apply dilation operation to joint
discrete areas. Since some defects lie in a vertical and
some others in a horizontal direction, we chose a
15×15 square structure element. The application of
these operations on the image in Figure 6 is shown in
Figure 7. The white regions in Figure 7 indicate
Figure 2. The background image of Figure 1 positions of defect in Figure 4. The defected region is
easily identified and separated with the use of this
image. The final result is shown in Figure 8.
The values of the structure elements scale and
threshold are selected by experience.
1072
Figure 4. A sample image of Sticking defect Figure 6. Result of converting Figure 5 to binary
image
1073
Figure 9. A sample image of Emulsion Rust Defect
5. Feature Extraction
Sample images of five defects are shown in Figure
4 and Figure 9-Figure 12. These defects have different
shapes and can be distinguished by their edges. So, we
should extract features that describe the difference
between neighborhood pixels. Appropriate features can
increase the accuracy of classification and bad features
decrease it. We chose nine features that describe these
defects. These features are as follows: entropy, Figure 10. A sample image of Under Pickled Defect
variance of the variances of the columns of the images
matrix, variance of variances of the rows of the images
matrix, median of image, average of variances of the
columns of the images matrix, average of variances of
the rows of the images matrix, average of local entropy
matrix, variance of averages of the columns of the
images matrix, power of correlation matrix of image.
All of these features are calculated over the detected
region of defect.
1074
selecting a suitable number of neurons of the hidden
layer, yet this number is important to determine the
accuracy of classification. If it is chosen low, the
network does not train well and its accuracy decreases.
If it is very high, the network adjusts itself with the
learning samples, hence the accuracy of classification
for the learning samples will increase but the accuracy
of classification for the test samples will decrease
[6,7].
In many cases, data does not strictly belong to one
class and because some classes are in close proximity
to each other, the data has a partial degree of
membership to each class. In this situation, Fuzzy
Inference System (FIS) is used to classify data. There
are two famous FIS’s methods, mamdani and sugeno.
Sugeno’s method uses fewer rules to describe data,
therefore it is faster than mamdani’s method for on-line
classification. We use this method in the simulation.
To generate FIS rules, there are several approaches;
Figure 11. A sample image of Dirty Surface Defect one of them is Fuzzy C-Mean (FCM) approach [8,9].
This algorithm classifies N feature
vectors X = {x1 , x2 ,..., x N } to C cluster by minimizing
Equation (1) which is called the cost function. This
minimization is done iteratively.
C n
¦¦ uijm vi − x j
2
J (U , v1 , v2 ,..vc ) = 1
i =1 j =1
1075
Step 3) Update the memberships uij in the was enhanced by special subtractive method and the
position of defect was determined using local entropy
membership matrix with Equation (4).
−1
and morphology. Finally, neural network classifies the
§ 2 · statistical feature vector extracted from the region of
¨ C § x j − vi · m −1 ¸
¦
uij = ¨ ¨ ¨ ¸
¨ l =1 ¨ x j − vl ¸¸
¸
¸
4
defect.
All of the used methods have acceptable accuracy
¨ © ¹ ¸ and time efficiency. By slight modification they can be
© ¹
Step 4) If the difference between the last two extended to cover more defects types. Also, because of
membership matrices is lower than į, or if the value of their low computation times, they can be employed in
the cost function is lower than Ȝ, terminate the online applications.
algorithm; otherwise, return to step 2.
9. Acknowledgement
7. Experiments and Comparison This work has been supported by Mobarakeh Steel
Company under grant No. 48222675.
We used 2300 labeled images from these five
defects to simulate the methods. 75 percent of the data 10. References
were chosen randomly to train the classifier and the
remaining 25% were used as test samples. The [1] Jong Pil Yun, YoungSu Park, Boyeul Seo,Sang Woo
accuracy of classification is defined as the ratio of the Kim, Se Ho Choi, Chang Hyun Park, Ho Mun Bae, Hwa
Won Hwang,” Development of Real-time Defect Detection
number of correct classification over the number of all
Algorithm for High speed Steel Bar in Coil(BIC)”, SICE-
data. ICASE International Joint Conference 2006, Oct. 18-2 1,
In our experiments, we used a 9×13×5 neural 2006 in Bexco, Busan, Korea.
network. The above experiment was repeated 1000 [2] Hongbin Jia, Yi Lu Murphey,Jianjun Shi, Tzyy-Shuh
times with random initialization values. In each Chang, “An Intelligent Real-time Vision System for Surface
simulation, the accuracy of classification was Defect Detection” , Proceedings of the 17th International
computed for the test and training data separately. The Conference on Pattern Recognition (ICPR’04).
average and variance were calculated and the results [3] N. Otsu, "A Threshold Selection Method from GrayLevel
are mentioned in Table 1. Histogram", IEEE Trans. Systems, Man and Cybernetics, Vo
l. 9, pp. 62-69, 1979.
[4] Rafael C.Gonzalez, Richard E.Woods, “Digital Image
Table 1. Accuracies of NN and FIS classification Processing”, Prentice Hall, New Jersey, 2008.
Accuracy of the Accuracy of the [5] J.Serra, “Image Analysis and Mathematical
training data test data Morphology”, Academic Press, New York, 1982.
Ave Var Ave Var [6] Martin T.Hagan, Howard B.Demuth, Mark Beale,
Feedforward “Neural Network Design”, Thomson Learning, Singapore,
97.89% 1.8×10-5 97.19% 5.1×10-5
NN 1996.
FIS- FCM [7] Richard O.Duda, Peter E.Hart, David G.Stork, “Pattern
82.47% 1.7×10-5 82.46% 1.3×10-5
method Classification”, John Wiley & Sons,China,2001.
[8]James C.Bezdek, “Pattern Recognition with Fuzzy
Considering the results of the above table, in which Objective Function Algorithms”, Plenum Press, New York,
the average of NN is greater than the average of FIS, 1981.
we chose F.F. NN as the classifier. Also, the small [9]James C.Bezdek, “Fuzzy Mathematics in Pattern
Classification”, Ph.D. thesis, Cornell University, Applied
value of variances shows that the network is valid. Mathematics Center, Ithaca, NY,1973.
[10] J.C. Bezdek, J. Keller, R. Krisnapuram, N.R. Pal, Fuzzy
8. Conclusion models and Algorithm for Pattern Recognition and Image
Processing, Kluwer Academic Publisher, 1999.
In this paper, we considered five famous surface
defects of cold rolling mill steel. The acquired image
1076