Measurement and Diagnosis of Concrete Cracks Using Image Processing and Fuzzy Theory

(1)

1 INTRODUCTION

A structure is required to be sustained during the target service life without any severe problems in safety, serviceability and durability. Most of rein- forced concrete structures, however, degraded serviceability and durability due to a very low tensile strength of concrete material even though it has quite high strength on compression. Especially, concrete structures with cracks due to tensile stresses show very fast degradation on serviceability and durability. Therefore, measuring and diagnosing cracks are essential to decrease direct or indirect effect of cracks on the safety, durability, and serviceability of concrete structures. In this regard, it is necessary to accurately measure the width, length, and direction of the cracks and deduce the causes of cracking based on the measuring data. However, because in- spectors generally measure cracks manually, a great deal of time and energy is required to take measurements and to compile relevant data.

This paper presents the crack measurement technique for detection and analysis of cracks in digital image of concrete surface to automate the measurement process of crack characteristics such as width, length, and orientation based on image processing to display the crack for inference and recognition of crack patterns, which includes horizontal, vertical, diagonal(-45°), diagonal(+45°), and random cracks,

based on artificial neural network. Diagnosis technique based on fuzzy theory also is presented.

2 METHODOLOGY 2.1 Crack measurement

Figure 1 shows the procedure of the proposed technique. The crack image taken with a digital camera contains 256 possible gray levels, ranging from 0 for black to 255 for white. At first, to correct for shade in an image, a morphology based on erosion, dilation, openness, and close operations were applied (Seul 2000). An image with a uniform background can be obtained by subtraction followed by openness, which consists of dilation followed by erosion.

After shade correction, in order to distinguish cracks from the background and from meaningless objects, an algorithm based on the binarization technique and shape analysis was developed. Figure 1 shows the algorithm. In a binary image, all pixels are assigned a value of 1 or 0 and these values represent an object that is based on a set threshold value. Determining the threshold value is therefore the most critical operation. Following a series of preliminary numerical tests, the discriminant method proposed by Otsu was adopted (Otsu 1979). This method was further im- proved by adding several crucial processes such as

Measurement and Diagnosis of Concrete Cracks Using Image Processing and Fuzzy Theory

B.Y. Lee, J.K. Kim & H. Myung

Department of Civil and Environmental Engineering, KAIST, Daejeon, Korea orea Railroad Corporation, Daejeon, Korea

ABSTRACT: This paper presents the crack measurement technique for detection and analysis of cracks in digital image of concrete surface. This technique composes of an automatic measurement process of crack characteristics such as width, length, and orientation based on image processing and inference and recognition process of crack patterns, which includes horizontal, vertical, diagonal(-45°), diagonal(+45°), and random cracks, based on artificial neural network. Diagnosis technique based on fuzzy theory also is presented. To verify the applicability of proposed technique, measurement and diagnosis program have been developed and a series of experimental investigations was carried out. A comparison of the original crack images with the results obtained by the proposed technique shows that the proposed technique can accurately detect and measure crack images. Presented crack pattern classifier based on ANNs can effectively classify cracks into 5 categories. Presented crack diagnosis program based on fuzzy theory presented same results with those by human and the possibility of each crack causes which is quantitative result.

Safety, Reliability and Risk of Structures, Infrastructures and Engineering Systems – Furuta, Frangopol & Shinozuka (eds)

(2)

noise reduction, local binarization, and a double extraction process.

The second extraction process aims to detect meaningful objects (cracks) that were undetected in the first extraction process. These objects resemble the background in terms of pixel values. They are also located out of the local binarization area, as the local binarization is performed according to the boundaries of the objects already detected by the global binarization. The second process uses the same binarization technique as the first extraction process but with a different image; in this image, the objects extracted during the first process have been completely eliminated.

Despite performing shade correction and binarization of the image, the cracks are still indistin- guishable from meaningless objects, mainly due to the failure to consider shape. Therefore, a modified shape factor as was considered a secondary criterion of crack detection. Initially a packing density index was adopted (Ammouche et al. 2000).

In order to calculate the features of detected cracks, a series of preprocessing steps was performed to obtain the thinned, boundary, and labeled (Haralick and Linda 1992) images. The crack width is defined as the sum of the minimum distances from the center pixel to both boundary pixels, and was calculated by applying the incorporated crack image to four distance filters. To calculate the crack length, after determining the starting point, depending on the location of the next pixel, the unit pixel length is

multiplied by 1.0 or 2, and this value is added to the previous value (the initial value was equal to ze- ro for the first step). By moving the position to the next labeled pixel, the iterative process continues until the end point is detected. The direction of the crack is simply calculated by using the coordinate relation between the starting point and the end point.

Real values are finally calculated by multiplying the values of the features obtained in the previous step by unit pixel length, which is determined in the step of image acquisition.

2.2 Crack pattern recognition

General process for pattern recognition is composed of preprocess, feature extraction, and recognition.

Preprocess is related with crack measurement. Fea- ture extraction is to characterize an object to be rec- ognized by measurements whose values are very similar for objects in the same category and very different for objects in different categories as well as to reduce the dimensions of the inputs. Recognition is to assign the object to a category based on the features provided (Richard et al. 2001).

To extract features, total projection technique is adopted in this study (Stroeven 1973). A feature is the projected length, which is normalized by maximum projected length to a range between 0 and 1, of cracks according to the rotation. Figure 2 shows the typical features for horizontal and random crack. Fi- nally, 60 features for each crack image are determined by averaging 180 projected lengths with 3 values. These features are used as inputs of classifier which is the artificial neural network in this study.

The capability of solving problems of artificial neural networks is determined by the complexity of the networks, which in turn may be determined by the number of layers, the complexity of neurons, the dynamic range of interconnections (weights and biases), and the number of hidden nodes. In general, the optimum architecture of an artificial neural network is realized by matching the complexity of the

Figure 1. Flow chart for crack measurement

(a) (b)

Figure 2. Features for (a) horizontal and (b) random cracks

(3)

neural network to the complexity of the problem.

That is, the architecture of the artificial neural network should be as simple as possible while retaining the capability of solving a given problem. The structure of networks is 60-4-5 (number of neurons in input layer, 60; number of neurons in hidden layer, 4;

and number of neurons in output layer, 5) (Figure 3).

Figure 4 shows the representative features for pattern recognition of cracks Table 1 represents the target values of output nodes for learning. Weights and biases are determined automatically by the training process. The Bayesian regularization (Mackay 1992, Foresee and Hagan 1997) is adopted as a learning algorithm in order to prevent over-fitting. The output node with maximum output value represents the relevant crack pattern.

Figure 5 shows the developed crack measurement and pattern recognition program

2.3 Crack diagnosis

It is difficult to deduce the causes of cracks quan- titatively because the causes of cracking are compli- cated and inter-related. Therefore, the fuzzy reasoning able to mimic the decision making process of an expert is adopted in this study (Zadeh 1965). The fuzzy theory was applied to the medical diagnosis system (Adlassnig and Kolarz 1982). General fuzzy inference process is composed of fuzzifying inputs, applying fuzzy operator, applying implication method, aggregating all outputs, and defuzzifying. The causes and symptoms presented by the JCI (1980) are used. The causes are classified into forty categories and symptoms are cracking time, location, pattern, conditions of placing, and mix-proportion.

Figure 6 shows the developed crack diagnosis program. The causes of cracks are deduced by three steps. First of all, the possibility values of cracking are calculated using the cracking time, type of cracking, and regularity and rules. And the second possibility values of cracking are also calculated using type of concrete deformation and range of cracking that is, materials, members, and structures. Lastly, the final possibility values of cracking are calculated on the first and second possibility values corrected by the conditions of placing, mix-proportion. Be- cause cracking time, temperature and humidity in placing, and unit cement content are continuous values, the rules which represent the degree of membership between those values and symptoms are constructed. Minimum operation is used for the calculation of possibility values of combination of

Figure 3. Structure of neural network

Table 1. Target values of output nodes in neural networks for learning

Horizon-

tal Vertical Diagonal (-45°)

Diagonal

(+45°) Random

node 1 1 0 0 0 0

node 2 0 1 0 0 0

node 3 0 0 1 0 0

node 4 0 0 0 1 0

node 5 0 0 0 0 1

Figure 5. Developed program for crack measurement

Figure 4. Representative features for pattern recognition of cracks

(4)

each symptom. On the other hand, maximum operation is used for the calculation of possibility values of crack causes from the possibility values of combination of each symptom. To diagnose conserva- tively the crack causes, minimum operation is also used for the calculation of possibility values of crack causes from the first and second possibility values of cracking.

Table 2 shows the rules for cracking time. I1, I2, and I3 mean one hour-one day, one day-28 days, and more than 28 days, respectively. Figures 7, 8, and 9 show the membership functions for temperature, humidity, and mix-proportion, respectively. And Figure 10 shows the membership functions for linguistic variable.

3 VERIFICATION AND DISCUSSION 3.1 Crack measurement

Several tests were performed on crack images taken with a digital camera to demonstrate the validity of the proposed algorithms. Figure 10 shows three crack images and the resulting detected crack images. Through a comparison of the original crack images with the results obtained by the proposed technique, the validity of the proposed technique is confirmed by the accurately detected crack images.

A test was performed on the crack width in order to verify the performance of the developed technique under the conditions at image acquisition. Table 3 provides comparisons between the crack widths calculated by the proposed technique and the values measured by an optical crack microscope (resolution: 0.1 mm). The results yielded very close numer-

Table 2. Rules for cracking time

time rule the time of cracking I1 I2 I3 1 very fast (0 0 1) VH L VL

2 fast (0 1 2) H L VL

3 slightly fast (1 2 4) SL H VL 4 normal (2 4 10) L VH L 5 slightly slow (4 10 90) VL NO NO 6 slow (10 90 365) VL L H 7 very slow (90 365 400) VL VL VH

* VH: very high, H: high, NO: normal, L: low, VL: very low Figure 6. Developed program for diagnosis

Figure 7. Membership function for temperature

Figure 8. Membership function for humidity

Figure 9. Membership function for mix-proportion

Figure 10. Membership function for linguistic variable

(5)

ical values. Table 4 and Table 5 show comparisons between the crack lengths (l_c) and crack directions (dc) calculated by the proposed technique and the values (l_m) measured by a ruler (resolution: 1 mm) and values (dm) measured by a protractor (resolution:

1°), respectively. It cannot be concluded that these results reflect the accuracy of the proposed technique owing simply to the errors in the measured data. However, at a minimum, the present results suggest that the proposed technique can provide suf- ficient accuracy for analyzing the widths of surface cracks in practical problems.

3.2 Crack pattern recognition

Table 6 shows the results of pattern recognition of cracks by the artificial neural network. It is exhibited that the classifier constructed in this study can effectively classify cracks into 5 categories.

3.3 Crack diagnosis

To verify crack diagnosis program, four examples were used. Table 6 shows the inputs for crack diagnosis and Table 7 shows the crack causes deduced by human and developed program, respectively. It can be seen that crack causes obtained using diagnosis program is the same with that deduced by a human. Furthermore, the developed program presents the possibility of each crack causes which is quantitative result.

4 CONCLUSION

This paper presented the crack measurement and diagnosis technique. To verify the applicability of proposed technique, measurement and diagnosis program have been developed and a series of experimental investigations was carried out. The following conclusions can be drawn from the current results:

(1) In order to enhance the accuracy of crack detection, a local binarization process and an extraction process were implemented. A comparison of the original crack images with the results obtained by the proposed technique shows that the proposed technique can accurately detect and measure crack images. Presented crack pattern classifier based on ANNs can effectively classify cracks into 5 categories.

(2) Presented crack diagnosis program based on fuzzy theory presented same results with those by human and the possibility of each crack causes which is quantitative result.

Figure 10. The results of crack detection tests Table 3. Comparison of crack width

Measurement point

1 2 3 4 5

wc: width calculated by proposed tech-

nique (mm)

1.05 1.51 0.85 0.86 1.02

wm: width measured by crack microscope

(mm)

1.0 1.5 0.8 0.9 1.0

Relative difference

(%) 5.5 0.67 5.9 4.7 2.3

Table 4. Comparison of crack length

Measurement point

1 2 3 4 5

lc: length calculated by proposed tech-

nique (mm)

680.6 312.3 614.4 561.5 112.6

lm: length measured

by ruler (mm) 630 300 580 540 100

Relative difference

(%) 8.03 4.10 5.93 3.98 12.6

Table 5. Comparison of crack direction

Measurement point

1 2 3 4 5

lc: length calculated by proposed tech-

nique (mm)

680.6 312.3 614.4 561.5 112.6

lm: length measured

by ruler (mm) 630 300 580 540 100

Relative difference

(%) 8.03 4.10 5.93 3.98 12.6

Table 6. Results of pattern recognition of cracks by artificial neural network

Classifications by artificial neural network Classifica-

tions by human

Hori- zontal

Ver-

tical -45° +45° Ran- dom Total

Accu- racy (%)

Horizontal 4 0 0 0 0 4 100

Vertical 0 19 0 0 0 19 100 Diagonal

(-45°) 0 0 7 0 0 7 100

Diagonal

(+45°) 0 0 0 5 0 5 100

Random 0 0 0 0 3 3 100

(6)

ACKNOWLEDGEMENTS

This study has been a part of a research project sup- ported by Korea Ministry of Education, Science and Technology (MEST) via the research group for control of crack in concrete. The authors wish to express their gratitude for the financial support that made this study possible.

REFERENCE

Adlassnig, K.P., Kolarz, G., 1982. CADIAG-2: Computer- Assisted Medical Diagnosis using Fuzzy Subsets (Edited by Gupta, M.M. and Sanchez, E., Approximate Reasoning in Decision Analysis), North Holland, pp.219-247.

Ammouche, A., Breysse, D., Hornain, H., Didry, O., and Mar- chand, J., 2000. A New Image Analysis Technique for The Quantitative Assessment of Microcracks in Cement-Based Materials, Cement and Concrete Research, 30(1), 25-35.

JCI, 1980. Guideline for crack investigation, repair/ streng- thening method.

Foresee, F.D. and Hagan, M.T., 1997. Gauss-Newton approx- imation to Bayesian Regularization, Proceedings of the 1997 International Joint Conference on Neural Networks, pp.1930-1935.

Haralick, R.M, and Linda, G.S., 1992. Computer and Robot Vision, Volume I, Addison-Wesley, pp. 28-48.

Mackay, D.J.C., 1992. Bayesian interpolation, Neural Compu- tation, 4(3), 415-447.

Otsu, N.A., 1979, Threshold Selection Method from Gray Lev- el Histogram, IEEE Transactions on Systems, SMC-9(1), 62-66.

Richard, O.D., Peter, E.H., and David. G.S., 2001. Pattern Classification, 2nd ed., John Wiley & Sons, Inc.

Seul, M., O'Gorman, L., and Sammon, MJ., 2000. Practical Algorithms for Image Analysis, Cambridge University Press.

Stroeven, P., 1973. Some Aspects of the Micromechanics of Concrete, Ph. D. Thesis, Stevin Laboratory, Technological University of DELFT.

Zadeh, L. A., 1965. Fuzzy sets, Information and Control. 8, 338-353.

Table 7. Comparison of crack causes by a human or program

Crack causes (human)

Program

Crack causes Possibility

1

B10 B4 D2 D4 D5

B10 19 0.68 B4

D2 D4 D5

13 35 37 38

0.6

2

A9 B2 B3

B2 11 0.58 B3 12 0.57 A9 9 0.53

3

A2 B2 B3 B14 B15

A2 2 0.8 B2 11 0.79 B3

B14 B15

12 23 24

0.7

4 A9

C1

C1 26 0.79 A9 9 0.7 Table 6. Inputs for crack diagnosis

Symptoms Input 1 2 3 4 Cracking time I1, I2, I3 30 21 365

Cement con-

tent I18, I19 340 340 350 340 Temperature I15, I16 32 18 33 30

Humidity I17 80 68 80 50

Net shape I4 0.2 0.2 0.2 0.2

Surface I5 0.3 0.7 0.7 0.3

piercing I6 0.8 0.9 0.5 0.9

Regularity I7 0.7 0.8 0.8 0.9 No regularity I8 0.3 0.2 0.2 0.1

Shrinkage I9 0.2 0.8 0.8 0.7

Swelling I10 0.2 0.2 0.2 0.8 Settlement,

flexural, shear I11 0.8 0.2 0.2 0.2 Material I12 0.5 0.5 0.5 0.3

Member I13 0.8 0.8 0.8 0.5

Structure I14 0.2 0.2 0.2 0.8