Author Affiliations
1Innovation Academy for Microsatellite, Chinese Academy of Sciences, Shanghai 201203, China2University of Chinese Academy of Sciences, Beijing 100049, Chinashow less
Fig. 1. YOLOv5s network structure
Fig. 2. Prediction process of YOLOv5
Fig. 3. Schematic diagram of Ghost module
Fig. 4. Schematic diagram of FA module
Fig. 5. FABottleneck design
Fig. 6. Flow chart of sparse parameter adaptive channel pruning algorithm
Fig. 7. Samples in DOTA dataset
Fig. 8. Image before and after segmentation. (a) Image before segmentation; (b) image after segmentation
Fig. 9. Comparison of detection effect before and after adding FA module. （a） Before adding FA module； （b） after adding FA module
Fig. 10. Gamma parameter distribution before pruning
Fig. 11. Gamma parameter distribution after pruning
| Module | YOLOv5s | YOLOv5m | YOLOv5l | YOLOv5x |
|---|
| depth_multiple | 0.33 | 0.67 | 1.0 | 1.33 | | width_multiple | 0.50 | 0.75 | 1.0 | 1.25 | | Number of C3 in backbone | 1，3，3 | 2，6，6 | 3，9，9 | 4，12，12 | | Number of C3 in neck | 1 | 2 | 3 | 4 | | Number of Conv | 32，64，128，256，512 | 48，96，192，384，768 | 64，128，256，512，1024 | 80，160，320，640，1280 |
|
Table 1. Number of YOLOv5 modules of different versions
| Model | Parameters /MB | Model size /MB | | Precision | Recall |
|---|
| YOLOv5s | 7.23 | 14.2 | 0.692 | 0.93 | 0.82 | | Ghost-YOLOv5s | 5.9 | 11.6 | 0.676 | 0.929 | 0.82 |
|
Table 2. Comparison between YOLOv5s and Ghost-YOLOv5s
| Type of targets | AP for different types of targets （YOLOv5s） | AP for different types of targets （Ghost-YOLOv5s） |
|---|
| All classes mAP | 0.692 | 0.676 | | plane | 0.942 | 0.939 | | baseball_diamond | 0.791 | 0.795 | | bridge | 0.543 | 0.518 | | ground_track_field | 0.653 | 0.589 | | small_vehicle | 0.600 | 0.586 | | large_vehicle | 0.806 | 0.799 | | ship | 0.878 | 0.871 | | tennis_court | 0.943 | 0.926 | | basketball_court | 0.683 | 0.657 | | storage_tank | 0.795 | 0.774 | | soccer_ball_field | 0.547 | 0.570 | | roundabout | 0.719 | 0.691 | | harbor | 0.819 | 0.825 | | swimming_pool | 0.711 | 0.718 | | helicopter | 0.638 | 0.532 | | container_crane | 0.012 | 0.019 |
|
Table 3. AP for different types of targets in YOLOv5s and Ghost-YOLOv5s
| Model | Parameters /MB | Model size /MB | | Precision | Recall |
|---|
| Ghost-YOLOv5s | 5.9 | 11.6 | 0.676 | 0.929 | 0.82 | | FA-YOLO | 3.75 | 7.64 | 0.673 | 0.947 | 0.79 |
|
Table 4. Comparison between Ghost-YOLOv5s and FA-YOLO
| Model | Parameters /MB | Model size /MB | | Precision | Recall |
|---|
| FA-YOLO （without FA） | 3.73 | 7.54 | 0.662 | 0.94 | 0.78 | | FA-YOLO （with FA） | 3.75 | 7.64 | 0.673 | 0.947 | 0.79 |
|
Table 5. Comparison before and after adding FA module
| Number of finetune | Pruning threshold | Parameters /MB | Model size /MB |
|---|
| 0.647 | 20 | 0.05 | 3.43 | 6.99 | | 0.646 | 20 | 0.1 | 3.18 | 6.49 | | 0.648 | 20 | 0.15 | 2.55 | 5.29 | | 0.577 | 20 | 0.2 | 1.41 | 3.11 | | 0.424 | 20 | 0.25 | 0.85 | 2.05 |
|
Table 6. Comparison of network models under different pruning thresholds
| Model | Precision | Recall | | Parameters /MB | Model size /MB | Mean inference time /s |
|---|
| YOLOv5s | 0.93 | 0.82 | 0.692 | 7.23 | 14.2 | 5.4 | | LW-YOLO | 0.943 | 0.79 | 0.648 | 2.55 | 5.29 | 5.2 |
|
Table 7. Comparison between Ghost-YOLOv5s and LW-YOLO
| Type of targets | AP for different types of targets |
|---|
| All classes mAP | 0.648 | | plane | 0.916 | | baseball_diamond | 0.712 | | bridge | 0.473 | | ground_track_field | 0.514 | | small_vehicle | 0.570 | | large_vehicle | 0.787 | | ship | 0.852 | | tennis_court | 0.933 | | basketball_court | 0.639 | | storage_tank | 0.738 | | soccer_ball_field | 0.470 | | roundabout | 0.615 | | harbor | 0.774 | | swimming_pool | 0.681 | | helicopter | 0.587 | | container_crane | 0.110 |
|
Table 8. LW-YOLO network AP and mAP for different types of targets
| Type of targets | Number of targets |
|---|
| plane | 8072 | | baseball_diamond | 412 | | bridge | 2075 | | ground_track_field | 331 | | small_vehicle | 126501 | | large_vehicle | 22218 | | ship | 32973 | | tennis_court | 2425 | | basketball_court | 529 | | storage_tank | 5346 | | soccer_ball_field | 338 | | roundabout | 437 | | harbor | 6016 | | swimming_pool | 2181 | | helicopter | 635 | | container_crane | 142 |
|
Table 9. Statistics of number of targets in each category in training set
