Chinese Journal of Ship Research, Volume. 20, Issue 3, 318(2025)
A detection algorithm for small surface floating objects based on improved YOLOv5s
Figures & Tables(14)
![[in Chinese]](/Images/icon/loading.gif){kind=icon}
Table 1. Presets for priori box sizes
View tableView in ArticleTable 1. Presets for priori box sizes
特征图 检测对象 先验框尺寸 160×160 极小目标 (5, 6) (8, 14) (15, 11) 80×80 小目标 (10,13) (16,30) (33,23) 40×40 中目标 (30,61) (62,45) (59,119) 20×20 大目标 (116,90) (156,198) (373,326)
Table 2. Comparison of detection layer pruning parameters
View tableView in ArticleTable 2. Comparison of detection layer pruning parameters
网络结构 mAP@0.5 mAP@0.95 Params GFLOP YOLOv5s 0.931 0.577 7012822 15.8 YOLOv5s+SmallTarget 0.941 0.586 7156008 18.5 YOLOv5s+SmallTarget_cut 0.951 0.615 2653590 15.9
Table 3. Loss function proportion table
View tableView in ArticleTable 3. Loss function proportion table
比例（IoU∶NWD） mAP@0.5 3∶7 0.946 4∶6 0.949 5∶5 0.955 6∶4 0.953 3∶7 0.951
Table 4. Comparison table of results for original dataset
View tableView in ArticleTable 4. Comparison table of results for original dataset
方法 mAP@0.5 mAP@0.95 YOLOv5s 0.897 0.469 YOLOv5s+SmallTarget_cut+ CBAM+NWD 0.916 0.484
Table 5. Partial experimental parameters
View tableView in ArticleTable 5. Partial experimental parameters
参数 数值 训练轮数epochs 300 批处理尺寸大小 batch size 16 学习率 0.01 优化器 SGD 动量 Momentum 0.937 输入图像大小 640×640
Table 6. Ablation test results
View tableView in ArticleTable 6. Ablation test results
实验编号 方法 mAP@0.5 mAP@0.95 Params GFLOP A YOLOv5s 0.931 0.577 7 012 822 15.8 B YOLOv5s+SmallTarget_cut 0.951 0.615 2 653 590 15.9 C YOLOv5s+CBAM 0.937 0.581 7 045 688 15.9 D YOLOv5s+NWD 0.937 0.587 7 012 822 15.8 E YOLOv5s+SmallTarget_cut+CBAM 0.954 0.62 2 664 636 15.9 F YOLOv5s+SmallTarget_cut+NWD 0.955 0.618 2 653 590 15.9 G YOLOv5s+NWD+CBAM 0.940 0.587 7 045 688 15.8 H YOLOv5s+SmallTarget_cut+CBAM+NWD 0.957 0.622 2 664 636 15.9
Table 7. Comparison of detection effects for different algorithms
View tableView in ArticleTable 7. Comparison of detection effects for different algorithms
网络结构 mAP@0.5 mAP@0.95 Params GFLOP YOLOv5s 0.931 0.577 7012822 15.8 YOLOv3-tiny 0.894 0.515 8666692 12.9 YOLOv5n 0.9 0.517 1760518 4.1 YOLOv8s 0.949 0.677 11135987 28.6 YOLOv5s+SmallTarget_cut+ CBAM+NWD 0.957 0.622 2664636 15.9
Table 8. Comparison of detection accuracy for different algorithms on the Visdrone2019 dataset
View tableView in ArticleTable 8. Comparison of detection accuracy for different algorithms on the Visdrone2019 dataset
方法 Pedestrain People Bicycle Car Van Truck Tricycle Awing-tricycle Bus Motor mAP@0.5 YOLOv5s 0.406 0.321 0.115 0.729 0.357 0.272 0.205 0.100 0.397 0.381 0.329 本文方法 0.411 0.328 0.113 0.740 0.359 0.312 0.240 0.126 0.483 0.403 0.351
Tools
Get Citation
Copy Citation Text
Xusheng YUE, Jun LI, Yaohong WANG, Penghao ZHU, Zhexing WANG, Xuanhao XU. A detection algorithm for small surface floating objects based on improved YOLOv5s[J]. Chinese Journal of Ship Research, 2025, 20(3): 318
Paper Information
Category: Weapon, Electronic and Information System
Received: Dec. 19, 2023
Accepted: --
Published Online: Jul. 15, 2025
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20