Optics and Precision Engineering, Volume. 31, Issue 2, 246(2023)
TCS-YOLO model for global oil storage tank inspection
Figures & Tables(21)
Fig. 1. Schematic diagram of each structure of the backbone layer of YOLOv5
Fig. 3. Ground-truth and multiple predicted bounding boxes for the same tank object
Fig. 9. Schematic diagram of oil storage tank target recognition by TCS-YOLO model
Table 1. Impact of C3TR layer replacement position on model performance
View tableView in ArticleTable 1. Impact of C3TR layer replacement position on model performance
Algorithm GFLOPs mAP0.5/% mAP0.5∶0.95/% YOLOv5 107.8 99.2 89.8 YOLOv5+C3TR
（position=1）
107.8 - - YOLOv5+C3TR （position=4） 104.9 99.3 91.1 YOLOv5+C3TR
（position=6，7，8）
99.0 99.1 89.9 YOLOv5+C3TR
（position=4，5，6，7，8）
93.1 99.1 89.4
Table 2. Effects of LN on C3TR
View tableView in ArticleTable 2. Effects of LN on C3TR
Model GFLOPs mAP0.5/% mAP0.5∶0.95/% YOLOv5 107.8 99.2 89.8 YOLOv5+C3TR
（with LN）
107.8 99.1 89.3 YOLOv5+C3TR （without LN） 104.9 99.3 91.1
Table 3. Effect of CBAM layer replacement position on model performance
View tableView in ArticleTable 3. Effect of CBAM layer replacement position on model performance
Algorithm GFLOPs mAP0.5/% mAP0.5∶0.95/% YOLOv5 107.8 99.2 89.8 YOLOv5+CBAM （position=1） 107.8 99.3 90.8 YOLOv5+CBAM
（position=4）
108.0 99.2 90.7 YOLOv5+CBAM
（position=1，4）
108.0 99.2 90.7 YOLOv5+CBAM
（position=1，2，3，4）
108.1 99.2 90.5 YOLOv5+CBAM
（position=5，6，7，8）
108.1 99.2 89.9 YOLOv5+CBAM
（position=1，2，3，4，5，6，7，8）
108.1 99.2 90.6
Table 4. Different IoU methods and results
View tableView in ArticleTable 4. Different IoU methods and results
IoU Method （A，G） （B，G） （C，G） （D，G） ℒIoU 0.3 0.3 0.6 0.6 ℒGIoU 0.3 0.3 0.6 0.6 ℒDIoU 0.3 0.3 0.624 0.624 ℒCIoU 0.3 0.3 0.624 0.624 ℒSIoU 0.306 0.31 0.721 0.862
Table 5. Influence of different IoU methods of localization loss functions on experimental results
View tableView in ArticleTable 5. Influence of different IoU methods of localization loss functions on experimental results
Algorithm mAP0.5
/%
mAP0.5∶0.95
/%
YOLOv5+IoU loss 99.1 89.6 YOLOv5+GIoU loss 99.2 89.8 YOLOv5+DIoU loss 99.2 89.8 YOLOv5+CIoU loss（default） 99.2 89.8 YOLOv5+SIoU loss 99.2 90.5
Table 6. Global oil storage tank dataset data source
View tableView in ArticleTable 6. Global oil storage tank dataset data source
Continent Country/City Company Number of images Africa
Algiers， Algeria Sonatrach 4 Tripoli， Libya Nation Oil Corporation 1 Cape Town， South Africa PetroSA 1 Asia Hong Kong， China Shell 1 Ningbo， China Sinopec 1 Xingang， China China National Petroleum Corporation 2 Tehran， Iran National Iranian Oil Company 1 Kawasaki， Japan Showa Shell Sekiyu 5 Osaka， Japan Eneos 4 Doha， Qatar Qatar Petroleum 1 Jeddah， Saudi Arabia Saudi Aramco 1 Singapore， Singapore Singapore Petroleum Company Limited 3 Abu Dhabi， UAE Abu Dhabi National Oil Company 6
Table 6. Global oil storage tank dataset data source
View tableView in ArticleTable 6. Global oil storage tank dataset data source
Continent Country/City Company Number of images Europe Den Haag， Netherland Royal Dutch Shell 2 North America Mexico City， Mexico Pemex 1 San Francisco， US Chevron Corporation 1 Houston， US Exxon Mobil 2 Oceania Brisbane， Australia Senex 1 Melbourne， Australia Oil Company of Australia 2 South America Riode Janeiro， Brazil Petrobras 3 Total - - 43
Table 7. Data set image parameter comparison
View tableView in ArticleTable 7. Data set image parameter comparison
Image Filming time Solar azimuth Satellite azimuth Cloudiness Abu Dhabi-A 2021-05-10 21：39：29 99.039 8 278.629 1% Abu Dhabi-B 2020-05-17 23：33：54 98.129 2 100.948 7%
Table 8. TCS-YOLO model training hyperparameter settings
View tableView in ArticleTable 8. TCS-YOLO model training hyperparameter settings
Hyperparameters Value Input resolution 416 Initial learning Rate 0.01 Training epochs 300 Momentum 0.937 Weight decay 0.000 5
Table 9. Batch-size value settings for different models
View tableView in ArticleTable 9. Batch-size value settings for different models
Models Batch-size TCS-YOLO-n 512 TCS-YOLO-s 256 TCS-YOLO-m 128 TCS-YOLO-l 64 TCS-YOLO-x 32
Table 10. Experimental results of different object detection models
View tableView in ArticleTable 10. Experimental results of different object detection models
Algorithm mAP0.5/% mAP0.5∶0.95/% Inference time/ms Parameters/M GFLOPs YOLOv3-tiny 96.4 77.5 0.5 33.1 12.9 YOLOv3 98.3 85.7 4.0 243.9 141.5 YOLOv3-spp 98.7 87.5 4.0 238.6 155.6 YOLOv4 97.3 84.7 3.7 234.6 154.7 YOLOv4-pacsp-mish-s 97.8 84.2 1.3 30.7 20.4 YOLOv4-pacsp-mish 98.1 86.0 4.3 200.3 119.3 YOLOv4-pacsp-mish-x 98.2 86.8 6.0 378.5 220.8 Swin_Transformer_mask_rcnn_tiny 94.4 74.2 1.9 326.65 540.5 Swin_Transformer_faster_rcnn 95.3 79.0 4.3 158.4 34.0 Swin_Transformer_mask_rcnn_small 96.8 82.0 5.6 408.0 555.7 YOLOv5n 99.0 84.8 0.3 6.7 4.2 YOLOv5s 99.0 88.4 0.8 26.9 16.3 YOLOv5m 99.1 89.7 1.9 80.3 50.3 YOLOv5l 99.2 89.8 3.1 177.8 114.1 YOLOv5x 99.2 89.9 5.3 332.7 217.1 TCS-YOLO-n 99.0 85.0 0.4 6.7 4.2 TCS-YOLO-s 99.3 89.4 0.9 26.8 15.7 TCS-YOLO-m 99.3 91.1 2.1 79.1 47.0 TCS-YOLO-l 99.4 91.3 3.4 175.8 111.2 TCS-YOLO-x 99.4 91.6 5.6 328.0 210.9
Table 11. Impact of different optimization methods on model performance
View tableView in ArticleTable 11. Impact of different optimization methods on model performance
Algorithm SIoU CBAM C3TR mAP0.5/% mAP0.5∶0.95/% YOLOv5 99.2 89.8 YOLOv5+SIoU loss √ 99.2 90.5 YOLOv5+CBAM √ 99.3 90.8 YOLOv5+C3TR √ 99.3 91.1 YOLOv5+SIoU loss+CBAM √ √ 99.3 91.2 YOLOv5+SIoU loss+C3TR √ √ 99.4 91.2 YOLOv5+CBAM+C3TR √ √ 99.4 91.2 TCS-YOLO （YOLOv5l+C3TR+CBAM+SIoU loss）√ √ √ 99.4 91.3
Tools
Get Citation
Copy Citation Text
Xiang LI, Rigen TE, Feng YI, Guocheng XU. TCS-YOLO model for global oil storage tank inspection[J]. Optics and Precision Engineering, 2023, 31(2): 246
Paper Information
Category: Information Sciences
Received: Jul. 15, 2022
Accepted: --
Published Online: Feb. 9, 2023
The Author Email: TE Rigen (terigen@jl1.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20