Laser & Optoelectronics Progress, Volume. 60, Issue 4, 0415006(2023)
Surface-Defect Detection Based on Feature Pyramid Matching and Self-Supervision
Figures & Tables(11)
Fig. 1. Frame diagram of defect detection based on feature pyramid matching and self-supervision
Fig. 6. Test results of some samples on MVtec AD dataset. (a) Original images; (b) ground truth; (c) abnormal score maps; (d) defective images
Table 1. Test results on MVTec AD dataset
View tableTable 1. Test results on MVTec AD dataset
Parameter Image AUROC / % Pixel AUROC /% SPADE KD-AD Proposed Proposed（SE） SPADE KD-AD Proposed Proposed（SE） Leather 94.47 95.26 100.00 100.00 96.94 98.15 99.24 99.53 Grid 75.36 78.52 94.62 97.25 93.72 91.78 99.00 98.16 Bottle 98.27 99.45 100.00 100.00 98.13 96.19 98.44 98.76 Transistor 82.53 85.47 95.96 94.42 93.98 76.04 89.26 89.18 Tile 88.68 91.63 98.99 100.00 87.47 82.31 96.53 95.95 Pill 80.24 83.30 94.93 97.77 96.44 90.08 96.02 97.71 Wood 91.39 94.51 99.04 99.72 88.26 84.65 95.90 96.37 Cable 85.42 89.38 98.20 97.23 97.05 82.59 97.38 97.70 Capsule 82.15 80.72 84.00 86.57 98.79 95.42 97.25 98.59 Carpet 77.04 80.05 97.03 95.04 96.82 95.74 98.98 99.22 Toothbrush 88.89 92.49 88.06 90.70 97.98 96.25 98.35 98.50 Zipper 90.12 93.58 95.01 94.67 96.48 94.24 97.27 98.03 Hazelnut 95.85 98.56 98.14 100.00 98.52 94.79 98.60 98.80 Metal_nut 71.94 73.74 89.15 98.15 98.21 86.25 92.41 98.30 Screw 80.86 83.95 92.29 90.54 98.81 95.84 96.79 98.90 Average 85.55 88.04 95.03 96.14 95.84 90.69 96.76 97.58
Table 2. Test results of other indicators on MVtec AD dataset
View tableTable 2. Test results of other indicators on MVtec AD dataset
Parameter Recall /% False rate /% Miss rate /% Time /ms SPADE 92.04 10.72 7.96 1530 KD-AD 91.93 7.62 8.07 84 Proposed 97.24 4.53 2.76 33 Proposed（SE） 97.68 3.84 2.32 35
Table 3. Test results on STL-10 dataset
View tableTable 3. Test results on STL-10 dataset
Parameter Image AUROC /% ResNet18 ResNet50 SE_ResNet50 ResNet18+BYOL ResNet50+BYOL SE_ResNet50+BYOL Truck 82.53 82.78 80.62 85.84 87.21 85.96 Horse 79.08 79.30 76.32 83.14 82.95 80.21 Deer 86.05 87.75 88.12 85.90 89.30 90.47 Car 86.93 89.35 89.93 90.38 92.48 92.78 Cat 73.16 73.89 76.89 77.75 77.52 79.53 Ship 92.38 93.47 92.97 93.74 94.08 93.62 Dog 72.45 73.81 72.78 69.84 70.94 68.48 Airplane 91.11 91.40 89.81 90.13 89.99 87.70 Monkey 79.73 82.69 83.86 80.20 83.57 84.62 Bird 77.09 79.23 78.47 77.85 80.15 78.85 Average 82.05 83.37 82.98 83.48 84.82 84.22
Table 4. Test results on CIFAR-10 dataset
View tableTable 4. Test results on CIFAR-10 dataset
Parameter Image AUROC /% 0 1 2 3 4 5 6 7 8 9 Average ResNet18 85.28 79.62 70.95 71.53 82.75 80.97 86.35 86.74 90.27 80.43 81.49 ResNet50 86.08 79.89 70.75 70.06 87.17 81.32 83.99 90.21 88.86 78.88 81.72 SE_ResNet50 83.45 77.25 69.86 73.95 89.45 77.34 87.72 89.82 85.38 76.96 81.12 ResNet18+resolution 86.69 81.57 70.57 73.18 87.56 80.05 88.71 85.86 89.02 81.69 82.49 ResNet50+resolution 86.56 80.92 70.48 73.35 88.29 79.97 86.08 87.99 88.24 80.14 82.20 SE_ResNet50+resolution 84.36 78.59 69.45 75.21 90.27 76.21 90.54 87.58 84.97 78.52 81.57 ResNet18+BYOL 85.38 81.65 68.19 73.26 87.28 83.37 87.36 91.32 90.73 83.14 83.17 ResNet50+BYOL 86.82 81.74 69.62 73.83 90.16 84.02 85.18 92.46 88.95 79.53 83.23 SE_ResNet50+BYOL 84.08 79.74 68.92 75.44 90.89 79.85 90.21 91.53 86.23 79.21 82.61
Table 5. Test results of other indicators on STL-10 and CIFAR-10 datasets
View tableTable 5. Test results of other indicators on STL-10 and CIFAR-10 datasets
Parameter STL-10 CIFAR-10 Recall /% False rate /% Miss rate /% Time /ms Recall /% False rate /% Miss rate /% Time /ms SE_ResNet50 91.21 14.35 8.79 34 90.02 15.58 9.98 34 SE_ResNet50+resolution 91.38 14.22 8.62 34 90.40 15.04 9.60 35 SE_ResNet50+BYOL 92.16 13.12 7.84 35 91.08 14.52 8.92 34 SE_ResNet50+resolution+BYOL 92.28 12.98 7.72 35 91.44 14.18 8.56 35
Tools
Get Citation
Copy Citation Text
Ming Liang, Minglu Zhang, Lü Xiaoling. Surface-Defect Detection Based on Feature Pyramid Matching and Self-Supervision[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0415006
Paper Information
Category: Machine Vision
Received: Nov. 11, 2021
Accepted: Dec. 21, 2021
Published Online: Feb. 14, 2023
The Author Email: Xiaoling Lü (lxl000418@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20