Laser & Optoelectronics Progress, Volume. 61, Issue 18, 1828004(2024)
Remote Sensing Object Detection Methods Based on Improved YOLOv5s
Figures & Tables(11)
Fig. 3. Structure of C3 and CA module. (a) Structure of C3 module; (b) structure of CA module
Fig. 4. Structure of DWCA and C3DWCA module. (a) Structure of DWCA module; (b) structure of C3DWCA module
Fig. 5. Structure of module. (a) Structure of FPN module; (b) structure of PANet module; (c) structure of BiFPN module
Table 1. Hyperparameters of network training
View tableTable 1. Hyperparameters of network training
Hyperparameter Value Learning rate 0.01 Learning rate decay 0.0001 Weight-decay 0.0005 Momentum 0.937 Batch 32 Epoch 200 Image size 800×800
Table 2. Results of ablation experiment
View tableTable 2. Results of ablation experiment
Model P R mAP@0.5 YOLOv5s 88.3 81.2 86.3 YOLOv5s+C3DWCA 94.5 91.3 95.4 YOLOv5s+BiFPN 89.9 86.9 91.6 YOLOv5s+C3DWCA+BiFPN 95.5 92.0 96.1
Table 3. Contrast of AP before and after modle improvement
View tableTable 3. Contrast of AP before and after modle improvement
Class AP /% Increment YOLOv5s Improved YOLOv5s Airplane 97.0 99.2 2.20 percentage points Airport 91.3 98.2 6.90 percentage points Baseball field 97.7 99.4 1.70 percentage points Basketball court 92.9 99.2 6.30 percentage points Bridge 59.0 83.7 24.70 percentage points Chimney 90.9 99.4 8.50 percentage points Dam 77.4 98.9 21.50 percentage points Expressway-service-area 94.1 99.3 5.20 percentage points Expressway-toll-station 91.7 97.4 5.70 percentage points Harbor 69.3 92.8 23.50 percentage points Golf course 78.2 98.8 20.60 percentage points Ground track field 92.8 98.5 5.70 percentage points Overpass 75.7 92.3 16.60 percentage points Ship 94.5 96.2 1.70 percentage points Stadium 93.2 98.7 5.50 percentage points Storage tank 88.6 92.0 3.40 percentage points Tennis court 96.3 98.9 2.60 percentage points Train station 77.6 98.6 21.00 percentage points Vehicle 72.5 82.7 10.20 percentage points Windmill 94.5 97.1 2.60 percentage points
Table 4. Contrast of model performance before and after modle improvement
View tableTable 4. Contrast of model performance before and after modle improvement
Model P /% R /% mAP /% FLOPs /G YOLOv5s 88.3 81.2 86.3 16.0 Improved YOLOv5s 95.5 92.0 96.1 17.5
Tools
Get Citation
Copy Citation Text
Kailun Cheng, Xiaobing Hu, Haijun Chen, Hu Li. Remote Sensing Object Detection Methods Based on Improved YOLOv5s[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1828004
Paper Information
Category: Remote Sensing and Sensors
Received: Jan. 11, 2024
Accepted: Feb. 5, 2024
Published Online: Sep. 9, 2024
The Author Email: Xiaobing Hu (scuhxb@163.com)
CSTR:32186.14.LOP240576
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20