Author Affiliations
1 College of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu 210016, China2 Wuxi Institute, Nanjing University of Aeronautics and Astronautics, Wuxi, Jiangsu 214187, Chinashow less
Fig. 1. Structural compositions of scratch defect detection system
Fig. 2. ROI images extracted under combination lighting mode. (a) Original image with low angle lighting; (b) original image with high angle lighting; (c) mask template; (d) ROI image
Fig. 3. Four kinds of morphological median filter kernels in different directions. (a) 0°; (b) 90°; (c) 45°; (d) 135°
Fig. 4. Background difference images for scratch extraction. (a) Background image by 13×13 median filter; (b) background image by morphological median filter; (c) background difference image by morphological median filter; (d) segmentation image of scratch binarization
Fig. 5. Schematic of scratch region growing
Fig. 6. Result of scratch region growing
Fig. 7. Flow chart of scratch detection based on weighted fusion of multi-features
Fig. 8. Scratch images extracted by different algorithms. (a) Gaussian filter; (b) median filter; (c) low-pass filter; (d) morphological median filter
Fig. 9. Scratch extraction error images obtained by different algorithms. (a) Gaussian filter; (b) median filter; (c) low-pass filter; (d) morphological median filter
Fig. 10. Experimental results of scratch detection
Characteristic parameter | Number of false inspections | Number of missed inspections | Correct rate |
---|
Area | 9 | 22 | 73.5% | Perimeter | 7 | 13 | 82.9% | Aspect ratio | 3 | 5 | 93.2% | Circularity | 4 | 5 | 92.3% | Rectangularity | 6 | 12 | 84.6% |
|
Table 1. Scratch detection accuracy based on single feature
Experimental algorithm | Image size /(pixel×pixel) | We/pixel | Er |
---|
Gaussian filter | 350×250 | 11279 | 0.1289 | Median filter | 350×250 | 4734 | 0.0541 | Low-pass filter | 350×250 | 15173 | 0.1734 | Morphology median filter | 350×250 | 1277 | 0.0146 |
|
Table 2. Scratch extraction errors under different algorithms
Detection method | Number of false inspections | Number of missed inspections | Correct rate | Detection time /s |
---|
Top-hat | 6 | 8 | 87.3% | 0.94 | Dual-threshold frequency domain differential | 3 | 5 | 90.8% | 1.61 | Proposed method | 2 | 5 | 95.7% | 1.21 |
|
Table 3. Scratch detection results under different methods