Defects Detection of Billet Surface Using Optimized Gabor Filters

Transcription

1 Proceedings o the 7th World Congress The International Federation o Automatic Control Deects Detection o Billet Surace Using Optimized Gabor Filters Jong Pil Yun* SungHoo Choi* Boyeul Seo* Chang Hyun Park* Sang Woo Kim* *Electronic and Electrical Engineering Department, Pohang University o Science and Technology, Pohang, Korea, ( {rebirth, csh45, boyeul80,chpark89, swkim}@postech.ac.kr) Abstract: This paper deals with a deects detection algorithm or billet surace. To get good perormance, detection algorithm has to solve several diicult problems such as shape o a billet, many kinds o deects, much scale. Especially, the scale, that is a metallic oxide on steel surace, makes worse severely the perormance o detection. To solve above problems, this paper presents a new eective deects detection algorithm based on Gabor ilters optimized by Genetic Algorithm. The experimental results conducted on real billet surace images and demonstrate the good perormance o the proposed algorithm. Keywords: Deect detection, Gabor ilter, Billet surace, Filter selection. INTRODUCTION In steel manuacturing industry, it becomes signiicant to enhance quality o products as well as to increase quantity o products. The deects on the surace o billet aect quality o BIC (Bar in Coil), thereore on-line surace inspection is o great importance to improve quality o steel products. However, the quality assurance o a billet surace is still carried out by human operators. This manual inspection cannot cover overall surace o a billet. Due to the dierent criterion o inspectors, the reliability and the accuracy o this inspection cannot be guaranteed (Yun et al., 006). So, in order to overcome manual inspection, many engineers concentrate on the automatic deects detection. There have been various approaches or the deects detection in many kinds o manuacturing such as ceramic tiles (Elbehiery et al., 005), rail suraces (Deutschl et al., 004), LCD(liquid Crystal Display) panel (Song et al., 006), textured materials (Kumar et al., 00 a), (Yang et al., 005), (Cho et al. 005), and so on. Although these methods have good perormance in the respective industries, it is hard to apply these methods to billet surace directly because each target has dierent eatures. Various approaches or deects detection have been proposed in steel manuacturing industry. The eed-orward neural network technique has been used to detect deects o cold rolled strips. The gray level arrangement o neighbor pixels was used to extract the eature vector (Kang et al., 005). A hybrid image segmentation method based on corner detection and Fisher discriminant has been presented to detect surace deects o ship plate (Guo et al., 006). The wavelet technique has been applied or deect detection o castings (Li et al., 006). In a hot rolling process, a real-time visual inspection system that used support vector machine to automatically learn complicated deect patterns has proposed (Jia et al., 004). In addition, many other methods applied to detection o steel surace deects (Sugimoto et al., 998), (Yun et al., 006), (Choi et al., 007). Although, these methods deals with steel products, it is hard to apply directly to billet deects detection since each methods are optimized on speciic steel type, lightening, and manuacturing speed. Moreover, the billet surace is covered with a lot o scale. Although it does not aect the quality o the product, it makes diicult to distinguish scale rom crack. In addition, scale o the billet has wide gray level range. These characteristics make it diicult to analysis image. So a new detection method which responds to only real deects, not scale is needed. In this paper, we propose a novel deects detection algorithm based on Gabor iltering method or solving the problem o detecting deects in billet surace. In order to obtain optimized Gabor ilters, GA (Genetic Algorithm) is used in this paper. And to remove noise still remained ater binarizing, the morphological operation is used. This paper is organized as ollows: Section analyzes intrinsic eatures o billet surace image. Section 3 presents deects detection algorithm based on the optimized Gabor ilters in detail. The perormance o the proposed deects detection algorithm is evaluated in Section 4. Finally, Section 5 gives conclusions.. ANALYSIS OF BILLET SURFACE IMAGE Fig. (a) shows billets surace image which is our target o deects detection. The billet is in the orm o long quadrangle stick. One o the our sides o billet is shown in Fig. (a). Fig. (b) shows gray level o a horizontal line rom A to B in Fig. (a). The gray level o both right and let sides is dierent rom center, because the billet has round corners not rectangular. There is a deect between 0-50 pixels o horizontal direction in Fig.. It is shown that the scale exists in not only deects parts but also all over the image. The scale is oxidized substance unavoidable caused during manuacturing process. The scale does not aect product quality, because it is eliminated during hot rolling process as a inished product process. But it makes diicult to detect deects, because the scale is mainly on the surace o billet /08/$ IFAC / KR

2 (a) (a) (b) (b) Fig.. Billet surace image with corner crack (a) original image (b) gray level o a horizontal line (c) st dierential value o a horizontal line (c) The light can relect o the scale part, which causes a very high gray level. On the other hand, it can also have a low gray level i the scale part comes out dark. I the scale has low gray level, it seems to be a deect. A deect exists between 40~50 pixels rom let in Fig.. As shown is Fig. (b), gray level o the deect is similar to scale. The gradient magnitude o gray level is calculated to analyze the image in greater detail. The gradient value o horizontal direction can be expressed like as ollows: D( x, y) = I( x, y) I( x, y) () where x and y are the spatial indies. I(x, y) is input image and D(x, y) is irst-order derivative. The gradient inormation using () is shown in Fig. (c) and Fig. (c). Not only deects but also scale has high gradient magnitude. This causes serious diiculty in the threshold level decision which discriminated between the deect the scale. Thereore, a new Fig.. Billet surace image with thin crack (a) original image (b) gray level o a horizontal line (c) st dierential value o a horizontal line algorithm or billet images must have ability to detect only deects among deects and scale. In this paper, we present the technique ocused on crack which has a huge eect on the product quality. The crack divided into two classes. As shown in Fig., the irst thing is located at right or let side o the image. The sizes are comparatively big, and the shapes are various. And a gray level o deects is low. We named this crack or corner crack. The corner crack is generated by rolling process, which is pre-processing o BIC process. The other thing is detected mainly in the middle part o an image, the width is very thin, typically ~ pixels. In this paper, this crack is called the thin crack. Because the gray level o thin crack is similar to that o scale, it is so diicult to distinguish crack rom scale. To solve above problems, we propose eective deects detection algorithm. In this paper, deects detection o billet surace using optimized Gabor ilters is investigated. Gabor ilters have been widely used in image processing and analysis. An (c) 78

3 important property o Gabor ilters is that they achieve maximum possible joint localization, or resolution in both spatial and spatial-requency domain. In spatial domain a Gabor unction is modulation product o complex exponential and a Gaussian envelope o arbitrary duration, in requency it is seen as a shited Gaussian (Shu et al., 004). Gabor ilters can decompose the image into components corresponding to dierent scales and orientation, thereore, have been used extensively or texture analysis, document analysis, object detection, and so on (Kumar., 00 a). The proposed deects detection algorithm o billet surace using optimized Gabor ilters is explained in next session. 3. PROPOSED DEFECTS DETECTION ALGORITHM 3. Gabor Function A two-dimensional Gabor unction consists o a sinusoidal plane wave o some requency and orientation, modulated by a two-dimensional Gaussian envelope. In order to conveniently ind a local spectral component o a texture, the real part o a Gabor unction is used (Mak et al., 005 a). In general, a two-dimension real part o a Gabor unction g(x, y) has the ollowing general orm: x y g ( x, y) = exp + cos( π x ) σ x σ y where x = xcosθ + ysinθ y = xsinθ + ycos θ, θ denotes the rotation parameter, is the radial requency o the Gabor unction. The space constants σ x and σ y deine the Gaussian envelope along the x and y axes. For a given input image I(x, y), the new image R(x, y) is obtained by using Gabor ilter g(x, y) as ollows: R( x, y) = g( x, y) I( x, y) M N = g ( mni, ) ( x my, n) m= 0 n= 0 where m, n are the Gabor ilter mask size variables, and (*) denotes the -D convolution. The squaring nonlinear operator computes the energy o every pixel in the iltered image R(x, y): E( x, y) = R ( x, y). (4) In this paper, the objective o Gabor ilter g(x,y) is to maximize the dierence between energy o the deect-ree billet surace and billet surace with deect. 3. Optimized Filter Selection () (3) The Gabor ilter has dierent response in the spatial and the requency domains according to our parameters such as σ x, σ y,, and θ. Also as the parameters o Gabor ilter change, the energy response E(x, y) o input image I(x, y) changes. Thereore the choice o Gabor ilter parameters is crucial role in the deects detection application. Our goal is to detect deects by enlarging the dierence o energy response between the deects and deect-ree. Thereore, two cost unctions using the energy mean are presented to ind the optimum parameters o Gabor ilter in this paper. And the optimized Gabor ilter is determined by genetic algorithm which searches maximum value o cost unction related energy separation criteria. The ollowing mean μ are used to represent the eature o the image. μ = E( x, y). (5) P Q ( x, y) region P and Q is size o given image. The irst cost unction is represented by ratio between the energy o deect image and deect-ree image with scale as ollows (Kumar et al., 006 b): J μ = (6) μ d Where μ d is the average energy o image with deect and μ is the average energy o deect-ree image with scale. The Gabor ilter g(x, y) that gives the highest cost unction J is chosen as the best eective ilter to detect deects. I Gabor ilter which is determined by using J apply to image with deects, the energy should be low, i.e., low μ d. On the contrary, the response o deects-ree image with scale should be strong, i.e., high μ. The second cost unction J is designated as normalized dierence o μ d and μ J μ μ = (7) d. μ In contrast to J, the second cost unction J increases along with μ d increases. On the other hand, the smaller μ is, the higher value o J becomes. In order to obtain optimized Gabor ilter which makes maximum value J or J, genetic algorithm is used in this paper. Because genetic algorithm searches the minimum value o the itness unction, the reciprocal o J is adopted. The chromosome vector is as ollows: C = [ σ x σ y θ] (8) The population size is 00. The selection, reproduction, crossover and random mutation are eiciently implemented. The mutation operation changes one o the parent based on a uniorm probability distribution. Because o constraints o computational simplicity, the size o Gabor masks is limited 79

4 to 9 9. Thereore, σ x and σ y is limited to The radial requency is limited 0-3 heuristically. Finally, θ lies between 0-π. 3.3 Thresholding and Binarizing The thresholding process discriminates between image with deects and deects-ree image with scale. It process is applied to R(x, y) passed through optimized Gabor ilter. In order to select optimal threshold value, a deect-ree image sample is used to generate a iltered image R(x, y). From this iltered image, the threshold value related to J is obtained as ollows: T = α min{ R( x, y) }, (9) x, y W where α is the weighting actor determined by the experiment. W is a window centered at the image R(x, y). The window size is chosen to avoid the possible undesirable eects due to border distortion (Kumar et al., 00 a). A binary algorithm is represented as ollows i T > R( x, y) then BI( i, j) = (0) else BI( x, y) = 0 where BI(x, y) is a binary image changed by the threshold value T. In the binary image, the value o means deects, and 0 means deect-ree. On the other hand, because J has maximum value at high μ d, T related to J has dierent value against T. The thresholding value T related to J is obtained as ollows: T = β max{ R( x, y) }, () x, y W where β is the weighting actor. And binary algorithm or T is represented as ollows i T < R( x, y) then BI( i, j) = else BI( x, y) = Morphological Operation () To remove small noise still remained ater binarizing, morphological technique is used. Erosion and dilation are the most basic morphological operators, which can serve as the undamental elements in the expressions o other complicated operators (Mak et al., 005 b). The dilation o A by B, denoted A B, is deined as A B = { z ( B ) z A }. (3) The dilation o A by B then is the set o all displacements, z, such that B and A overlap by at least one element (Gonzalez et al., 00). The relection o set is denoted B * is deined as * B w w b b B = { =, or }. (4) The translation o set A by point z=(z +z ), denoted (A) z, is deined as ( A) = { c c = a + z, or a A}. (5) z The erosion o A by B, denoted A B, is deined as A B = { z ( B) z A}. (6) In words, this equation indicates that the erosion o A by B is the set o all points z such that B, translated by z, is contained in A (Gonzalez et al., 00). When an image is eroded or dilated by a structuring element, some inormation in the original image will be lost. In order to recover the lost inormation caused by erosion, dilation and to remove small noise in the B(x, y), opening operation and closing operation are adopted. The opening o set A by structuring element B, denoted A B, is deined as A B = ( A B) B. (7) And the closing operation is deined as ollow: A B = ( A B) B. (8) Thus, the opening A by B is the erosion o A by B, ollowed by a dilation o the result by B. On the other hand, the closing operation is a contrary concept o the opening operation. The structuring element B inluences the elimination rates o noise. Thereore, selection o the structuring element B is very important. And also, the size o structuring element B inluences the calculation speed. Hence, a structuring element is decided on by the maximum size o a noise and minimum size o a deect (Cho et al., 005). Based on the analysis results o corner crack and thin crack, the opening operation mask B =[ ] T and the closing operation mask B =[ ] T are used or corner crack. The opening operation mask C =[ ] T and the closing operation mask C =[ ] T are applied to thin crack. 4. EXPERIMENT AND RESULTS To evaluate the perormance o the proposed detection algorithm, we used the billet images directly acquired rom the real production line. The size o images is 04 by 000. The parameters o optimized Gabor ilters are determined by the (6) and (7) or the corner crack and thin crack respectively. To search maximum value o the each cost unction (6) and (7), the genetic algorithm is used. In the experiment results, Equation (6) is suitable or separating deect region and deect-ree region or billet image with corner crack. On the contrary, (7) is it or thin crack. The parameters o the Gabor ilters optimized by genetic algorithm is shown in Table. Fig. 3(a) and (b) illustrate the optimized Gabor ilters or corner crack and thin crack respectively. The perormance o the proposed detection algorithm is evaluated by using 370 real billet images in which 80

5 (a) (b) Fig. 3. Optimized gabor ilters (a) or corner crack (3 9) (b) or thin crack (5 5) Table. Gabor ilter parameters obtained by the optimized ilter selection method σ x σ y θ (radian) Corner crack Thin crack Fig. 4. Example or detection results o corner crack (a) 9 images are deect-ree and the rest is with deects. The test results are summarized in Table. The proposed algorithm has 9.3% and 83.83% accuracy or corner crack and thin crack respectively. Some o the detection results or the billet image are presented in Fig Fig. 4-6 are the results or the image with corner crack and Fig. 7 show the results or thin crack. On the other hand, Fig. 8 shows the results or a billet image with scale. Each (b) in Fig. 4-8 are images passed through the optimized Gabor ilters and each (c) is binary image obtained ater thresholding operation. Final results ater noise removal using morphological operation are shown in Fig 4-8 (d). Especially as shown Fig. 8 (d), the scale which seems to be deects is not detected. Consequently, the proposed deects detection algorithm using optimized Gabor ilters is eective or billets surace image with scale. Fig. 5. Example or detection results o corner crack (a) 5. CONCLUSIONS In this paper, a new deect detection algorithm based on the optimized Gabor ilters or billet surace is presented. The billet surace image has a lot o scale which is oxidized substance unavoidably caused during manuacturing process and extensions. Because scale o billet surace is similar to real deects such as crack, it is diicult to distinguish deects rom image with scale. To solve this problem, we propose optimization approaches utilizing energy separation criteria. The optimized Gabor ilter is determined by genetic algorithm which searches maximum value o cost unction related energy separation criteria. The image, which is passed through the optimized Gabor ilters, changes to binary image using thresholding. To remove small noise still remained ater binarizing, morphological technique is used. In this paper, two structuring element ilters proposed or each deects. To evaluate the perormance o the proposed algorithm, the detection eiciency is measured by experiment using the real billet image. Experiment results show that the proposed algorithm is eective and suitable or billet image with scale. Fig. 6. Example 3 or detection results o corner crack (a) 8

6 Fig. 7. Example or detection results o thin crack (a) Graylevel billet image (b) Filtered image (c) Binary image Fig. 8. Example o detection results o image with scale (a) Table. The experiment results o the proposed detection algorithm Deects Non-deects success ail success ail accuracy Corner crack % Thin crack % REFERENCES C. -S. Cho, B. -M. Chung, and M. -J. Park (005). Development o Real-Time Vision-Based Fabric Inspection System. IEEE Trans. Industrial Electronics, 5 No. 4, S. H. Choi, J. P. Yun, B. Seo, Y. S. Park, and S. W. Kim (007). Real-time Deects Detection Algorithm or Highspeed Steel Bar in Coil. Enormatika Trans. on Engineering, Computing and Technology, 9, E. Deutschl, C. Gasser, A. Niel, and J. Werschonig (004). Deect Detection on Rail Suraces by a Vision based System. IEEE Intelligent Vehicles Symposium, H. Elbehiery, A. Henawy, and M. Elewa (005). Surace Deects Detection or Ceramic Tiles Using Image Processing and Morphological Techniques. IEEE Trans. Engineering, Computing and Technology, 5, R. C. Gonzalez, R. E. Woods (00). Digital Image Processing. Chapter 9, Prentive-Hall, Upper Saddle River, New Jersey. J. H. Guo, X. D. Meng, and M. D. Xiong (006). Study on Deection Segmentation or Steel Surace Image Based on Image Corner Detection and Fisher Discriminant. Journal o Physics: Con. Series 48, H. Jia, Y. L. Murphey, J. Shi, and T. S. Chang (004). An Intelligent Real-time Vision System or Surace Deect Detection. Proceedings o the International conerence on Pattern Recognition., 3, G.- W. Kang, H. -B. Liu (005). Surace Deects Inspection o Cold Rolled Strips Based on Neural Network. Proc. International Conerence on Machine Learning and Cybernetics, 8, A. Kumar, G. K. H. Pang (00)a. Deect Detection in Textured Materials Using Gabor Filters. IEEE Trans. Industry Applications, 38 No., A. Kumar, G. K. H. Pang (00)b. Deect Detection in Textured Materials Using Optimized Filters. IEEE Trans. System, Man, Cybernetics, 3 No. 5, X. Li, S. K. Tso, X.-P. Guan, Q. Huang (006). Improving Automatic Detection o Deects in Castings by Applying Wavelet Technique. IEEE Trans. Industrial Electronics, 53 No. 6, K. L. Mak, P. Peng, and H. Y. K. Lau (005)a. A Real- time Computer Vision System or Detecting Deects in Textile Fabrics. IEEE International Conerence on Industrial Technology, K. L. Mak, P. Peng, H.Y,K. Lau (005)b. Optimal Morphological Filter Design or Fabric Deect Detection. IEEE International Conerence on Industrial Technology, Y. Shu, and Z. Tan (004). Fabric Deects Automatic Detection Using Gabor Filters. Proc. World Congress on Intelligent Control and Automation, 4, Y.- C. Song, D.- H. Choi, K.- H. Park (006). Wavelet-Based Image Enhancement or Deect Detection in Thin Flim Transistor Liquid Crystal Display Panel. Japanese Journal o Applied Physics, 45 No. 6A, T. Sugimoto, T. Kawaguchi (998). Development o a Surace Deect Inspection System Using Radiant Light rom Steel Products in a Hot Rolling Line. IEEE Trans. Instrumentation and Measurement, 47 No., X. Yang, G. Pang, and N. Yung (005). Robust Fabric deect detection and classiication using multiple adaptive wavelets. IEE Proc.-Vision, Image and Signal Processing, 5 No. 6, J. P. Yun, Y. S. Park, B. Seo, S. W. Kim, S. H. Choi, C. H. Park, H. M. Bae, and H. W. Hwang (006). Development o Real-time Deect Detection Algorithm or High-speed Steel Bar in Coil(BIC). SICE-ICASE International Joint Conerence,