Optics and Precision Engineering, Volume. 32, Issue 11, 1746(2024)
Worm surface defect detection with fusion of multi-scale features
Figures & Tables(18)
Fig. 5. Distribution of label sizes in worm surface defect training dataset
Fig. 7. Impact of different neck networks on recall for various categories
Fig. 9. Effectiveness of different algorithms on worm gear surface defect detection
Table 1. Worm shooting camera and lens parameters
View tableView in ArticleTable 1. Worm shooting camera and lens parameters
Property Value Resolution 2 448×2 048 Pixel size/ 3.45 Working distance/mm 73 Frame rate/（frame·s-1） 36 Field of view/mm2 Magnification 0.438 Unit pixel represents physical distance/ 7.89
Table 2. Distribution of worm raw data defect
View tableView in ArticleTable 2. Distribution of worm raw data defect
Defect type Defect code Defect quantity Ii 1 1 609 Sc 2 414 Fc 3 252 Sd 4 669
Table 3. Distribution of surface defect data in worm dataset
View tableView in ArticleTable 3. Distribution of surface defect data in worm dataset
Defect type Train Val Test Total Ii 1 548 242 205 1 995 Sc 1 372 140 144 1 656 Fc 804 108 96 1 008 Sd 1 674 171 144 1 989 Image quantity 3 862 480 463 4 805
Table 4. Experimental results adding different neck networks to model
View tableView in ArticleTable 4. Experimental results adding different neck networks to model
Neck Parameters/M P/% R/% mAP@0.5/% FPS PAN 36.49 80.89 68.4 73.6 60.24 SPPAFPN 35.81 76.12 71.8 73.9 65.79 AFPN 28.73 76.03 74.5 75.5 75.75
Table 5. Experimental results of inserting EMA at different positions
View tableView in ArticleTable 5. Experimental results of inserting EMA at different positions
Project code Insertion position Parameters mAP@0.5/% - - 28.73 75.5 1 4，5 28.75 75.9 2 6，7 28.75 75.6 3 8，9，10 28.79 75.6 4 2，3 29.61 74.7 5 1，2 28.94 76.5
Table 6. Experimental results using different loss functions
View tableView in ArticleTable 6. Experimental results using different loss functions
Index Loss function mAP@0.5 mAP@0.75 1 CIoU 76.5 26.5 2 DIoU 74.4 25.5 3 GIoU 73.7 25.2 4 EIoU 73.4 24.8 5 MPDIoU 74.2 26.4 6 WIoUv1 74.4 25.3 7 WIoUv2 75.3 26.7 8 WIoUv3 74.8 24.6 9 SIoU 76.9 25.3
Table 7. Ablation experiments of three improvement approaches
View tableView in ArticleTable 7. Ablation experiments of three improvement approaches
AFPN EMA SIoU Paramete/M mAP@0.5 Variation 0 0 0 36.49 73.6 - 1 0 0 28.7 75.5 +1.9 0 1 0 36.5 74.3 +0.7 0 0 1 36.49 73.8 +0.2 1 1 0 28.94 76.5 +2.9 0 1 1 36.51 75.1 +1.5 1 0 1 28.7 74.5 +0.9 1 1 1 28.94 76.9 +3.3
Table 8. Test results of different algorithms on worm gear surface defect dataset
View tableView in ArticleTable 8. Test results of different algorithms on worm gear surface defect dataset
Algorithm Image Size Parameter/M mAP@0.5 mAP@0.75 YOLOv3 640×640 61.51 69.9 18.1 YOLOR-p6 640×640 36.82 45.9 9.5 YOLOv5m 640×640 20.8 73.3 - YOLOv7 640×640 36.49 73.6 23.1 YOLOv8m 640×640 25.84 72.2 35.5 AFP-YOLO 640×640 28.94 76.9 25.3
Table 9. Actual detection performance of AFP-YOLO among 30 worms
View tableView in ArticleTable 9. Actual detection performance of AFP-YOLO among 30 worms
Defect Defect
quantity
Detect
quantity
False detect Miss detect Ii 316 312 1 5 Sc 137 131 2 8 Fc 135 133 1 3 Sd 229 229 2 2 Total 817 805 6 18
Tools
Get Citation
Copy Citation Text
Lei WANG, Wenping GUO, Xinwei CHEN, Min XIA. Worm surface defect detection with fusion of multi-scale features[J]. Optics and Precision Engineering, 2024, 32(11): 1746
Paper Information
Category:
Received: Jan. 8, 2024
Accepted: --
Published Online: Aug. 8, 2024
The Author Email: GUO Wenping (wpguo@hust.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20