3. Results and discussion
3.1. Experimental equipment
Special lighting angles and backgrounds are beneficial to segment the defects from the magnetic tile image. In order to accomplish the defects detection task of different curvature radii in all surfaces, we designed an automatic defects detection system for magnetic tiles, and we can adjust the position and angle of the light and camera programmatically. The system schematic diagram is demonstrated in Fig. 6.
Full-size image (24 K)
Fig. 6.
System schematic diagram.
Figure options
The experimental camera is an array CCD camera with external triggering. It can move along the linear guide to adjust the distance between the camera and the magnetic tile, which allows to easily adjust the depth of focus programmatically for different kinds of magnetic tiles. The linear guide rail installed on the arc guide rail can make arc adjustment to meet the requirement of different radians of magnetic tiles. The light source uses white LEDs, whose brightness and switch are adjustable. The light board installed on the arc guide rail can also make arc adjustment to meet the requirement of different locations of light in different stations for magnetic tile crack defects detection.
Fig. 7 shows the system prototype. The conveyor belt on the system prototype is driven by a servo motor. The speed of the magnetic tile transportation can be adjusted to test detection efficiency on the experimental equipment and to simulate the actual means of transport in production. The optical fiber sensor triggers the camera when conveyor belt deliveries the magnetic tiles to the detected position. Images taken by camera are put into the computer processing and control system via image collection card. Then the images can be used for defects detection.
Full-size image (48 K)
Fig. 7.
The system prototype.
Figure options
3.2. Crack extraction
We used the FDCT to decompose Fig. 1(a). The eigenvalues of the texture feature measurement were calculated in reconstructed images with different k values, as shown in Fig. 8. It is evident from Fig. 8(b) that the best choice of k is between 0.01 and 0.02, and then the segmentation threshold T was obtained according to Eq. (11).
Full-size image (47 K)
Fig. 8.
Eigenvalues of the texture feature measurement in restructured images: (a) 0≤k≤0.9 and (b) 0≤k≤0.1.
Figure options
Reconstructed images with different k values are shown in Fig. 9, where the mask method was used. It can be seen that the image almost loses all its details when k is quite small, such as k=0.001; and when k equals some suitable value, such as k=0.01, regular texture is virtually eliminated; as the value of k increases, the texture barely changes. In Fig. 9, Fig. 9(a) almost loses all the details; Fig. 9(b and c) eliminates the regular textures and largely retain the crack defects; Fig. 9(d–f) almost recovers to original images.
Full-size image (24 K)
Fig. 9.
Restructured images based on different remaining coefficients: (a) k=0.001, (b) k=0.005, (c) k=0.01, (d) k=0.05, (e) k=0.1, and (f) k=0.5.
Figure options
The relationship between the curvelet transform coefficients and the reconstrcted image is nonlinear because of the textures themselves in the image.
We used Canny operator to extract contours of Fig. 9, the results are shown in Fig. 10. Crack defects cannot be extracted when k=0.001; when k=0.005 and k=0.01, the results of the crack defects detection are acceptable; when k is larger than 0.05, it difficult to distinguish the defects because the cracks are mixed with the grinding textures. The experiment results verify that the method of obtaining k value introduced in this paper is correct and valid.
Full-size image (36 K)
Fig. 10.
Extracted contours of Fig. 9 by Canny operator: (a) k=0.001, (b) k=0.005, (c) k=0.01, (d) k=0.05, (e) k=0.1, and (f) k=0.5.
Figure options
It should be noted that the requirements of the light and background are different for different surfaces on the same magnetic tile. Reason for this is that by adjusting the light and background, the cracks can be shown most clearly. Thresholds for reconstructed coefficients are also different for different surfaces on the same magnetic tile. Therefore the thresholds need to be obtained through tests in advance.
According to the above method, we extracted the crack in Fig. 1(b), the result is shown in Fig. 11. As you can see, the arc texture in the outside arc surface has been eliminated and only a vertical crack left. We did the mask to Fig. 11 and used the morphology operation to eliminate the chamfered contour, and then the crack defect can be segmented completely.
Full-size image (9 K)
Fig. 11.
Extracted crack defect in the outside arc surface.
Figure options
In order to verify the effect of this method and show the superiority of the algorithm, we selected four different magnetic tiles with defects in different surfaces to extract cracks. The extraction results are compared with the Otsu algorithm and the morphological filtering algorithm in Fig. 12. As we can see, the Otsu algorithm and the morphological filtering algorithm cannot extract crack defects accurately because of the intensity inhomogeneity and low contrast of the magnetic tile images while the method in this paper can effectively detect the cracks.
Full-size image (38 K)
Fig. 12.
Extracted cracks in different surfaces: (a) original images, (b) detection results by Otsu algorithm, (c) detection results by morphological filtering algorithm, and (d) detection results by the proposed method.
Figure options
3.3. Judging defect
In actual production processing, cracks shorter than 1 mm in the end face or shorter than 3 mm in the outside arc surface are not considered to be defects, so we should find a way to determine whether the crack is a defect or not. First, the algorithm applies the dilation operation based on mathematical morphology to the reconstructed image in order to connect the adjacent line segments. Second, the erosion operation also based on mathematical morphology is used to restore the width of the extracted contour, then we can calculate the length, the area and the Hu moments of the extracted contour of crack by Canny operator. The ratio of the area to the length can give a preliminary judge whether the crack is a defect or not. Generally speaking, the defect crack's width is less than 10 pixels while its length contains dozens of pixels. So the ratio of the area to the length can be set according to the statistics in actual testing. The crack is considered to be a defect when the actual value is larger than the preset value, or, it is a qualified product. In order to increase the accuracy of the decision, we calculate the Hu moments of the extracted contour of crack to judge whether the extracted contour is a crack.
Fig. 13 shows the efficiency of our method for small cracks. The small cracks in the end face are shown in Fig. 13(a), and the result of extracted cracks is shown in Fig. 13(b). The actual length of the smallest crack is 0.8 mm by measuring all of them. It illustrates that the method proposed in this paper can be used to detect small cracks in actual production.
Full-size image (20 K)
Fig. 13.
Extracted cracks in the end face: (a) magnetic tile with small cracks and (b) extracted contour of Fig. 13(a).
Figure options
3.4. Results and discussion
Multiple experimental results suggest that cracks shorter than 0.8 mm could not be detected by the proposed method. In the actual production process, however, a crack shorter than 1 mm is usually not considered to be a defect, so the method in this paper can meet the industrial detection requirement.
In this experiment, the image size is 659×494 pixels, and the length of the magnetic tile is 70 mm. We adjust the camera position so that the length of the magnetic tile is approximately the width of the image, such that each pixel represents about 0.1 mm in length. But it is only the theoretical value by using the sub-pixel segmentation technology. The actual detection accuracy of the crack is lower because of the regular grinding texture. Since the period of the grinding texture is not a fixed value which is generally between 0.05 mm and 0.1 mm, the theoretical length of the cracks can be extracted is greater than 0.1 mm.
In order to test our method, we selected a batch of magnetic tiles of 59 defective products and 56 qualified products. Test results are shown in Table 1. The accuracy rate is 93.9%, and the false acceptance rate is 3.57%. Reason for missing the cracks is that the reflection from some magnetic tile surfaces masks the cracks. The correct rejection rate is 8.5%, which demonstrates that a small percentage of qualified magnetic tiles are misidentified because of some deep grinding textures.
3. Results and discussion
3.1. Experimental equipment
Special lighting angles and backgrounds are beneficial to segment the defects from the magnetic tile image. In order to accomplish the defects detection task of different curvature radii in all surfaces, we designed an automatic defects detection system for magnetic tiles, and we can adjust the position and angle of the light and camera programmatically. The system schematic diagram is demonstrated in Fig. 6.
Full-size image (24 K)
Fig. 6.
System schematic diagram.
Figure options
The experimental camera is an array CCD camera with external triggering. It can move along the linear guide to adjust the distance between the camera and the magnetic tile, which allows to easily adjust the depth of focus programmatically for different kinds of magnetic tiles. The linear guide rail installed on the arc guide rail can make arc adjustment to meet the requirement of different radians of magnetic tiles. The light source uses white LEDs, whose brightness and switch are adjustable. The light board installed on the arc guide rail can also make arc adjustment to meet the requirement of different locations of light in different stations for magnetic tile crack defects detection.
Fig. 7 shows the system prototype. The conveyor belt on the system prototype is driven by a servo motor. The speed of the magnetic tile transportation can be adjusted to test detection efficiency on the experimental equipment and to simulate the actual means of transport in production. The optical fiber sensor triggers the camera when conveyor belt deliveries the magnetic tiles to the detected position. Images taken by camera are put into the computer processing and control system via image collection card. Then the images can be used for defects detection.
Full-size image (48 K)
Fig. 7.
The system prototype.
Figure options
3.2. Crack extraction
We used the FDCT to decompose Fig. 1(a). The eigenvalues of the texture feature measurement were calculated in reconstructed images with different k values, as shown in Fig. 8. It is evident from Fig. 8(b) that the best choice of k is between 0.01 and 0.02, and then the segmentation threshold T was obtained according to Eq. (11).
Full-size image (47 K)
Fig. 8.
Eigenvalues of the texture feature measurement in restructured images: (a) 0≤k≤0.9 and (b) 0≤k≤0.1.
Figure options
Reconstructed images with different k values are shown in Fig. 9, where the mask method was used. It can be seen that the image almost loses all its details when k is quite small, such as k=0.001; and when k equals some suitable value, such as k=0.01, regular texture is virtually eliminated; as the value of k increases, the texture barely changes. In Fig. 9, Fig. 9(a) almost loses all the details; Fig. 9(b and c) eliminates the regular textures and largely retain the crack defects; Fig. 9(d–f) almost recovers to original images.
Full-size image (24 K)
Fig. 9.
Restructured images based on different remaining coefficients: (a) k=0.001, (b) k=0.005, (c) k=0.01, (d) k=0.05, (e) k=0.1, and (f) k=0.5.
Figure options
The relationship between the curvelet transform coefficients and the reconstrcted image is nonlinear because of the textures themselves in the image.
We used Canny operator to extract contours of Fig. 9, the results are shown in Fig. 10. Crack defects cannot be extracted when k=0.001; when k=0.005 and k=0.01, the results of the crack defects detection are acceptable; when k is larger than 0.05, it difficult to distinguish the defects because the cracks are mixed with the grinding textures. The experiment results verify that the method of obtaining k value introduced in this paper is correct and valid.
Full-size image (36 K)
Fig. 10.
Extracted contours of Fig. 9 by Canny operator: (a) k=0.001, (b) k=0.005, (c) k=0.01, (d) k=0.05, (e) k=0.1, and (f) k=0.5.
Figure options
It should be noted that the requirements of the light and background are different for different surfaces on the same magnetic tile. Reason for this is that by adjusting the light and background, the cracks can be shown most clearly. Thresholds for reconstructed coefficients are also different for different surfaces on the same magnetic tile. Therefore the thresholds need to be obtained through tests in advance.
According to the above method, we extracted the crack in Fig. 1(b), the result is shown in Fig. 11. As you can see, the arc texture in the outside arc surface has been eliminated and only a vertical crack left. We did the mask to Fig. 11 and used the morphology operation to eliminate the chamfered contour, and then the crack defect can be segmented completely.
Full-size image (9 K)
Fig. 11.
Extracted crack defect in the outside arc surface.
Figure options
In order to verify the effect of this method and show the superiority of the algorithm, we selected four different magnetic tiles with defects in different surfaces to extract cracks. The extraction results are compared with the Otsu algorithm and the morphological filtering algorithm in Fig. 12. As we can see, the Otsu algorithm and the morphological filtering algorithm cannot extract crack defects accurately because of the intensity inhomogeneity and low contrast of the magnetic tile images while the method in this paper can effectively detect the cracks.
Full-size image (38 K)
Fig. 12.
Extracted cracks in different surfaces: (a) original images, (b) detection results by Otsu algorithm, (c) detection results by morphological filtering algorithm, and (d) detection results by the proposed method.
Figure options
3.3. Judging defect
In actual production processing, cracks shorter than 1 mm in the end face or shorter than 3 mm in the outside arc surface are not considered to be defects, so we should find a way to determine whether the crack is a defect or not. First, the algorithm applies the dilation operation based on mathematical morphology to the reconstructed image in order to connect the adjacent line segments. Second, the erosion operation also based on mathematical morphology is used to restore the width of the extracted contour, then we can calculate the length, the area and the Hu moments of the extracted contour of crack by Canny operator. The ratio of the area to the length can give a preliminary judge whether the crack is a defect or not. Generally speaking, the defect crack's width is less than 10 pixels while its length contains dozens of pixels. So the ratio of the area to the length can be set according to the statistics in actual testing. The crack is considered to be a defect when the actual value is larger than the preset value, or, it is a qualified product. In order to increase the accuracy of the decision, we calculate the Hu moments of the extracted contour of crack to judge whether the extracted contour is a crack.
Fig. 13 shows the efficiency of our method for small cracks. The small cracks in the end face are shown in Fig. 13(a), and the result of extracted cracks is shown in Fig. 13(b). The actual length of the smallest crack is 0.8 mm by measuring all of them. It illustrates that the method proposed in this paper can be used to detect small cracks in actual production.
Full-size image (20 K)
Fig. 13.
Extracted cracks in the end face: (a) magnetic tile with small cracks and (b) extracted contour of Fig. 13(a).
Figure options
3.4. Results and discussion
Multiple experimental results suggest that cracks shorter than 0.8 mm could not be detected by the proposed method. In the actual production process, however, a crack shorter than 1 mm is usually not considered to be a defect, so the method in this paper can meet the industrial detection requirement.
In this experiment, the image size is 659×494 pixels, and the length of the magnetic tile is 70 mm. We adjust the camera position so that the length of the magnetic tile is approximately the width of the image, such that each pixel represents about 0.1 mm in length. But it is only the theoretical value by using the sub-pixel segmentation technology. The actual detection accuracy of the crack is lower because of the regular grinding texture. Since the period of the grinding texture is not a fixed value which is generally between 0.05 mm and 0.1 mm, the theoretical length of the cracks can be extracted is greater than 0.1 mm.
In order to test our method, we selected a batch of magnetic tiles of 59 defective products and 56 qualified products. Test results are shown in Table 1. The accuracy rate is 93.9%, and the false acceptance rate is 3.57%. Reason for missing the cracks is that the reflection from some magnetic tile surfaces masks the cracks. The correct rejection rate is 8.5%, which demonstrates that a small percentage of qualified magnetic tiles are misidentified because of some deep grinding textures.
การแปล กรุณารอสักครู่..