From: Review of vision-based steel surface inspection systems
Paper | Method | Type of defects | Sample size | Features | Classification accuracy (%) | Resolution (Across x along) (mm) | Speed of steel object | Real-time operation | |
---|---|---|---|---|---|---|---|---|---|
Detection | Classification | ||||||||
[46] - slab | Robert edge filter, zonal thresholding, edge connectivity | Statistical tree classifier | Crack, pits, tear, scab, rollmark, under fill, bleeder |  | 17 geometric |  | 0.076 × 1.27 | 4 in./s (36.6 m/min) | Suitable |
[70] - hot strip | Five wavelets (feature extraction) | VVRKFA, SVM | 24 types (details in paper) | 1,432 images | 3 to 9 elements - 5 wavelets | 93.8 | 1 × 1 | 20 m/s |  |
[36] - hot strip | Co-occurrence matrix | NN-BP, GMDH, SOM | RIS, bruise, rolled in bruise, scratch, lamination | 1,084 cases | 24 spatial, textural | NN-BP 83.5 (ave.) | Â | Â | Â |
[62] - hot strip | FFT (feature extraction), GA | LVQ | Crack, pits, scar, rollmark, shell, cross | 485 images | 54 frequency domain | 84 to 93 | Â | Â | Â |
texture, pseudo | |||||||||
[32] - hot strip | Â | NN-BP | Shell, hole, pit/scab, residue scale, coil break, slip mark, rust | 255 samples | 4 geometric | 98.75 | Â | 10 to 15Â m/s | Â |
[31] - cold strip | FFT (filtering), local entropy, thresholding, morphology | NN-BP, fuzzy logic | Sticking, emulsion rust, under pickled, dirty surface, anneal stain | 2,300 images | 9 statistical | NN-BP: 97, Fuzzy logic: 82 | 0.37 | Â | Suitable |
[76] - cold strip | Sobel edge detector, region growing | Fuzzy logic | Scaling, shrunken leaf, cusping, file mark, cleavage crack, tear | 15 to 20 defects/class | Geometric, fractal | 85 to 95 | Â | 6Â m/s | Suitable |
[26] - cold strip | Thresholding (six types) | Decision tree discrimination logic | Scab, sliver, scale, gauge, scratch, roll mark, oil spot | 196 defects |  | 95.5 | 0.17 × 1.25 | 1,400 m/min (23 m/s) | Suitable |
[89] - cold strip | Â | LVQ | Rust, lamination, rollmark, scratch, weld, and stain | 135 images | 17 feature +11 reference | Approximately 68 to 96 | Â | 5Â m/s | Â |
[63] - cold strip | Multifractal | NN-BP | Sticking, emulsion rust, under pickled, dirty surface, anneal stain defect | 2,300 images | 10 geometric, gray level | 97.9 | 0.37 | Â | Â |
[43] - cold strip | Morphological, thresholding | Hough transform: line defect | Welding, clamp, hole, oxidation, waveform, exfoliation | 300 images | 6 components | Line defect: 98. | Â | Â | Â |
PCA-SOM: complex shape | Complex shape: 77 | ||||||||
[87] - cold strip | Sobel edge detector, threshold, special filter | Fuzzy logic | Large population of inclusions | 212 images | 4 | 95 | Â | Â | Â |
[83] - cold strip | Difference image | SVM | Line scab, pickle patch, hole, slip mark, oil drop, pit | 1,200 images | 54 geometric, gray level | 92.4 | Â | Â | Â |
[82] - cold strip | Â | Weak classifier | Serrated edge, scratch, shell, inclusion, weld, wrinkles. | 500 images | Extended Haar rectangle features | 94 | Â | Â | Â |
[77] - cold strip | Â | Modified NN BP | Flash, roller moulage, oxidation skin, hole, crack | 300 | Â | 94.34 | Â | Â | Â |
[59] - rod/bar | Special horizontal, vertical, diagonal edge filters | Hierarchical | Seam, scratch, roll mark, overfill | 663 images | 12 geometric, gray level | NN BP: 90.66, RVM: 91.21 | Â | 15Â m/s | Suitable |
NN BP, RVM | |||||||||
[85] - rod/bar |  | Process knowledge-based multiclass SVM (PK-MSVM) | Seams, longitudinal cracks, scales, transverse mid-surface cracks and transverse corner cracks | 600 bars (20,140 defects) | 4 process features + image data set | 81.4 to 93.9 |  |  |  |