Object Detection Method Based on Improved YOLOv4 Network for Remote Sensing Images

Zhenjiu Xiao; Yueying Yang; Xiangxu Kong

doi:10.3788/LOP213399

Laser & Optoelectronics Progress, Volume. 60, Issue 6, 0628009(2023)

Object Detection Method Based on Improved YOLOv4 Network for Remote Sensing Images

Zhenjiu Xiao, Yueying Yang^*, and Xiangxu Kong

Author Affiliations

College of Software, Liaoning Technical University, Huludao 125105, Liaoning, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Target detection network structure of improved YOLOv4

Download full size

Fig. 2. Operating process of Pyramid MaxpoolNMS

Download full size

Fig. 3. Principle of Mosaic data enhancement

Download full size

Fig. 4. P-R curve of oil tank in improved YOLOv4 algorithm category

Download full size

Fig. 5. Loss trend during model training

Download full size

Fig. 6. Comparison of test results before and after improvement. (a) YOLOv4 algorithm; (b) improved YOLOv4 algorithm

Download full size

Fig. 7. AP and mAP of improved YOLOv4 algorithm on RSOD dataset

Download full size

Table 1. Dataset category

View table

Table 1. Dataset category

Class

plane

ship

tank

tennis court

basketball

court

baseball

court

track-and-field ground

bridge

port

vehicle

Training set

1290

661

1283

1076

204

370

165

142

302

1020

Test set

555

284

550

462

159

131

437

Total

1845

945

1833

1538

293

529

239

204

433

1457

Table 2. Performance comparison of different improvement methods
View table
Table 2. Performance comparison of different improvement methods
Method MobileNetV3 Self-attention PSRR-MaxpoolNMS mAP /% FPS
YOLOv4 87.17 33
1 √ 85.16 57
2 √ 87.67 29
3 √ 86.19 36
4 √ √ 86.81 49
5 √ √ 85.62 59
6 √ √ 87.23 35
Proposed method √ √ √ 86.50 54

Table 3. Comparison of parameters, model size, and FPS of different methods
View table
Table 3. Comparison of parameters, model size, and FPS of different methods
Method Parameter /10⁶ Model size /MB FPS
Faster R-CNN 76.81 7
SSD512 62.54 102.50 18
YOLOv3 63.95 246.50 19
YOLOv4 61.13 244.30 33
YOLOv5s 7.20 35.70 63
Propoed Method 11.70 44.70 54

Table 4. Comparison of different detection methods in each category

View table

Table 4. Comparison of different detection methods in each category

Method	Faster R-CNN	SSD	YOLOv3	YOLOv4	YOLOv5s	Propoed method
mAP	80.96	75.80	84.90	87.17	81.90	86.50
plane	0.94	0.84	0.99	0.99	0.95	0.99
ship	0.82	0.62	0.84	0.82	0.76	0.87
tank	0.65	0.78	0.89	0.88	0.86	0.87
tennis court	0.82	0.79	0.99	0.95	0.88	0.99
basketball court	0.89	0.88	0.60	0.81	0.70	0.75
baseball court	0.95	0.89	0.96	0.95	0.90	0.97
track-and-field ground	0.92	0.80	0.98	0.94	0.92	0.97
bridge	0.58	0.65	0.62	0.66	0.71	0.64
port	0.72	0.71	0.91	0.87	0.84	0.89
vehicle	0.77	0.62	0.70	0.81	0.67	0.71

Tools

Get Citation

Copy Citation Text

Zhenjiu Xiao, Yueying Yang, Xiangxu Kong. Object Detection Method Based on Improved YOLOv4 Network for Remote Sensing Images[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0628009

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites