Optics and Precision Engineering, Volume. 33, Issue 9, 1496(2025)
Improvement of YOLOv8 for multi-scale defect detection in wind turbine blades
Figures & Tables(21)
Fig. 3. Structure layout of DualConv （
Fig. 12. Comparison of mAP curves before and after algorithm improvement for wind turbine blade defects
Fig. 14. Visualization comparison before and after YOLOv8 model improvement
Fig. 15. Comparison of mAP curves before and after algorithm improvement for PCB defects
Table 1. Ablation Experiment
View tableView in ArticleTable 1. Ablation Experiment
Number 160×160
detection layer
Improved
neck
ACFM GRE-SPPF DE-C2f P（%） R（%） mAP@0.5（%） mAP@0.5∶0.95（%） Params/M FPS 1 — — — — — 86.8 77.2 84.9 55.4 3.01 111.1 2 √ — — — — 88.2 79.5 86.9 57.3 2.92 104.2 3 — √ — — — 85.5 81.6 87.6 59.1 2.94 107.5 4 — — √ — — 86.3 78.5 85.6 55.1 3.51 90.1 5 — — — √ — 83.7 80.4 85.8 55.5 3.07 108.7 6 — — — — √ 85.6 82.2 88.4 59.1 2.66 86.2 7 — √ — — √ 88.0 80.7 88.5 59.9 2.58 75.2 8 — √ √ — — 84.1 83.5 88.0 58.4 3.46 84.0 9 — √ √ √ — 86.5 83.9 88.8 59.6 3.50 79.4 10 — √ — √ √ 85.8 83.4 88.8 61.7 2.62 74.6 11 — √ √ √ √ 88.6 84.9 91.1 61.8 3.16 63.7
Table 2. Comparison of effectiveness of attention mechanism modules
View tableView in ArticleTable 2. Comparison of effectiveness of attention mechanism modules
Model P（%） R（%） mAP@0.5（%） mAP@0.5∶0.95（%） Params/M M3+ACFM 84.1 83.5 88.0 58.4 3.46 +CBAM ［ 19 ］88.3 78.8 87.6 58.1 3.03 +EMA 84.1 82.7 87.5 57.7 2.94 +MCA ［ 20 ］88.6 77.5 86.9 57.0 2.94 +SE ［ 21 ］87.0 81.1 87.2 59.1 2.99 +ECA 84.4 81.7 87.3 57.5 2.94 +GAM ［ 22 ］83.3 82.3 88.0 59.2 5.12 +iEMA ［ 23 ］84.7 83.5 87.2 58.3 3.04 +SEAM ［ 24 ］87.6 81.7 87.6 59.2 3.05 +MLCA ［ 25 ］87.3 81.9 87.7 58.4 2.94
Table 3. Comparison of effectiveness of different SPPF improvement modules
View tableView in ArticleTable 3. Comparison of effectiveness of different SPPF improvement modules
Module P（%） R（%） mAP@0.5（%） mAP@0.5：0.95（%） Params/M SPPF 88.5 83.3 89.2 61.8 3.10 SimSPPF［ 26 ］88.6 81.9 89.6 62.6 3.10 SPPFCSPC［ 27 ］88.9 81.5 88.9 60.7 4.71 SPPF-LSKA［ 28 ］89.6 79.1 88.9 60.3 3.37 DPAM［ 29 ］87.9 80.9 89.7 61.2 3.20 SPPFAPGC［ 30 ］85.8 82.0 87.0 57.8 3.24 GRF-SPPF［ 31 ］86.6 84.9 89.6 61.5 3.17 SPPF-S［ 32 ］86.8 83.5 89.7 60.7 3.11 GRE-SPPF 88.6 84.9 91.1 61.8 3.16
Table 4. Comparison of model performance with DE-C2f module at different locations
View tableView in ArticleTable 4. Comparison of model performance with DE-C2f module at different locations
Location P（%） R（%） mAP@0.5（%） mAP@0.5∶0.95（%） Params/M B2-B5 87.8 80.3 87.8 58.3 3.22 B2-B5&P3-P4&N2-N5 89.6 84.3 90.0 61.2 2.92 B2-B5&N2-N5 86.9 79.5 87.1 58.1 2.97 B2-B5&P3-P4 88.2 82.8 89.1 59.7 3.16 P3-P4&N2-N5 85.9 81.7 88.1 59.5 3.22 N2-N5 86.2 82.7 87.6 58.9 3.28 B3-B5&N2-N5 87.3 81.8 88.5 58.4 2.98 B3-B5&P3-P4&N3-N4 88.9 82.9 90.3 60.0 3.11 B3-B5&P3-P4 88.6 84.9 91.1 61.8 3.16
Table 5. Model comparison experiment
View tableView in ArticleTable 5. Model comparison experiment
Model P（%） R（%） mAP@0.5（%） mAP@0.5∶0.95（%） Params/M FPS Faster-RCNN 30.8 65.9 52.0 25.0 137.10 - SSD 36.1 62.3 53.4 27.3 26.29 - RT-DETR-l 77.8 69.1 76.6 45.6 31.99 63.2 YOLOv5n 87.6 76.3 84.6 55.3 2.50 108.7 YOLOv6 80.4 75.2 79.8 49.8 4.23 125.0 YOLOv7-tiny 84.5 74.4 81.0 48.3 6.02 - YOLOv8n 86.8 77.2 84.9 55.4 3.01 111.1 YOLOv8s 88.0 76.9 85.5 56.9 11.10 116.3 YOLOv9s 86.5 78.6 86.6 59.8 7.17 64.5 YOLOv10n 82.0 75.9 82.9 53.3 2.70 104.2 YOLOv11n 84.6 79.1 85.7 55.1 2.58 66.2 Ours 88.6 84.9 91.1 61.8 3.16 63.7
Table 6. Comparison of experimental results on PCB defect dataset
View tableView in ArticleTable 6. Comparison of experimental results on PCB defect dataset
Model AP@0.5 mAP@0.5 mAP@0.5∶0.95 Missing hole Mouse bite Open circuit Short Spur Spurious copper YOLOv8
Ours
99.5
99.5
88.8
90.9
89.9
94.8
96.7
96.6
83.2
89.9
94.1
93.0
92.1
94.1
46.7
51.6
Tools
Get Citation
Copy Citation Text
Guang ZHU, Chen GU, Liyun XU, Yanqiong SHI, Zhengyang DING, Xu ZHANG, Yonghua ZHANG. Improvement of YOLOv8 for multi-scale defect detection in wind turbine blades[J]. Optics and Precision Engineering, 2025, 33(9): 1496
Paper Information
Category:
Received: Dec. 16, 2024
Accepted: --
Published Online: Jul. 22, 2025
The Author Email: Guang ZHU (guangzhu123@ahjzu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20