Remote Sensing Technology and Application, Volume. 39, Issue 3, 590(2024)
Object Detection in Remote Sensing Images based on YOLOX-Tiny Biased Feature Fusion Network
Figures & Tables(17)
Fig. 11. MAP50 radar chart of different models under different categories
Table 1.
Classes of different objects and the number of View tableView in ArticleTable 1.
Classes of different objects and the number of Train Val Trainval Test C1 344 338 682 705 C2 326 327 653 657 C3 551 577 1 128 1 312 C4 336 329 665 704 C5 379 495 874 1 302 C6 202 204 406 448 C7 238 246 484 502 C8 279 281 560 565 C9 285 299 584 634 C10 216 239 455 491 C11 536 454 990 1 322 C12 328 332 660 814 C13 410 510 920 1 099 C14 650 652 1 302 1 400 C15 289 292 851 619 C16 391 384 775 839 C17 605 630 1 235 1 347 C18 244 249 493 501 C19 1 556 1 558 3 114 3 306 C20 404 403 807 809 Total 5 862 5 863 11 725 11 738
Table 2.
The ablation experiment results of our model on the DIOR dataset View tableView in ArticleTable 2.
The ablation experiment results of our model on the DIOR dataset 方法 Params
/M
P/% R/% APS/% APM /% APL/% AP/% FPS
/f/s
mAP50 /% ①YOLOX-Tiny 5.04 85.87 63.43 11.3 35.3 62.7 43.9 47 69.79 ②YOLOX-Tiny+MBTFPN 4.89 86.30 64.14 12.1 37.0 63.6 45.5 49 71.15 ③YOLOX-Tiny+MBTFPN+ Deformable convolution 4.99 88.18 65.19 12.7 38.9 69.6 49.5 46 73.02 ④YOLOX-Tiny+MBTFPN+ deformable convolution +SIoU 4.99 88.24 65.59 12.8 39.0 69.8 49.6 46 73.68
Table 3.
The Comparative experiment results of different models on the DIOR dataset View tableView in ArticleTable 3.
The Comparative experiment results of different models on the DIOR dataset Method Neck Size Params/M GFLOPs/G FPS/f/s mAP50/% Faster R-CNN FPN 640 28.5 948.4 6 63.10% CenterNet / 640 32.7 109.3 19 56.05% YOLOv4-Tiny SPP+PAN 640 5.9 16.2 74 61.01% YOLOv5-s PAN 640 7.1 16.5 50 66.97% YOLOv7-Tiny SPPCSPC+优化的PAN 640 6.1 13.3 66 73.43% Our optimized model MBTFPN 640 4.9 15.8 46 73.68%
Table 4.
Detection performance of different models for targets with different scales on DIOR dataset View tableView in ArticleTable 4.
Detection performance of different models for targets with different scales on DIOR dataset Method Neck Size APS/% APM /% APL/% Faster R-CNN FPN 640 6.5 32.3 57.6 CenterNet / 640 5.4 25.2 51.4 YOLOv4-Tiny SPP+PAN 640 6.7 31.3 50.5 YOLOv5-s PAN 640 11.1 37.4 62.0 YOLOv7-Tiny SPPCSPC+优化的PAN 640 12.3 38.9 70.1 Our optimized model MBTFPN 640 12.8 39.0 69.8
Table 5.
The Comparative experiment results of different models on the RSOD datasets View tableView in ArticleTable 5.
The Comparative experiment results of different models on the RSOD datasets Method Size Params/M FPS/f/s aircraft overpass oiltank playground mAP50/% Faster R-CNN 640 28.5 6 71.90% 100% 90.90% 100% 90.70 CenterNet 640 32.7 19 70.83% 85.37% 86.56% 100% 85.69 YOLOv4-Tiny 640 5.9 74 78.28% 84.58% 83.94% 100% 86.70 YOLOv5-s 640 7.1 50 94.10% 80.50% 95.10% 99.10% 92.20 MDCF2Det 640 - - 81.02% 100% 90.77% 100% 92.95 YOLOX-Tiny 640 5.04 47 95.95% 86.93% 96.64% 100% 94.88 Our optimized model 640 4.9 46 97.34% 91.59% 99.55% 100% 97.12
Tools
Get Citation
Copy Citation Text
Zhaohua HU, Yuhui LI. Object Detection in Remote Sensing Images based on YOLOX-Tiny Biased Feature Fusion Network[J]. Remote Sensing Technology and Application, 2024, 39(3): 590
Paper Information
Category:
Received: Nov. 22, 2022
Accepted: --
Published Online: Dec. 9, 2024
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20