Author Affiliations
School of Electronics and Information Engineering, Hebei University of Technology, Tianjin, 300401, Chinashow less
Fig. 1. YOLOv5 network structure
Fig. 2. Long edge definition method
Fig. 3. PSANeck structure
Fig. 4. PSA module structure
Fig. 5. ECALayer structure
Fig. 6. YOLOv5 PAN structure
Fig. 7. Improved feature fusion structure
Fig. 8. Densely coded label
Fig. 9. Improved network structure diagram
Fig. 10. YOLOv5m multi-scale characteristic thermodynamic diagram. (a) High resolution branch; (b) medium resolution branch; (c) low resolution branch
Fig. 11. Improved algorithm multiscale characteristic thermodynamic diagram. (a) High resolution branch; (b) medium resolution branch; (c) low resolution branch
Fig. 12. Comparison of detection effect between YOLOv5m and improved algorithm. (a)(b)(c) detection results of YOLOv5m algorithm; (d) (e) (f) detection results of improved algorithm
| PL | SH | SV | LV | mAP |
---|
180/8 | 86.55 | 86.90 | 73.58 | 76.82 | 80.96 | 180/32 | 87.45 | 87.33 | 76.00 | 77.40 | 82.04 | 180/64 | 87.46 | 87.34 | 76.00 | 80.20 | 82.75 | 180/128 | 87.44 | 87.32 | 76.01 | 80.18 | 82.73 | 180/180 | 87.39 | 87.33 | 76.04 | 79.90 | 82.67 |
|
Table 1. Comparison of model performance under different angle discretization granularity
Group | PSA | ECA | BiFPN | DCL | PL | SH | SV | LV | GFlops | mAP |
---|
G1 | × | × | × | × | 78.79 | 79.46 | 64.61 | 61.23 | 53.7 | 71.02 | G2 | √ | × | × | × | 86.53 | 79.47 | 64.60 | 69.34 | 49.8 | 74.99 | G3 | √ | √ | × | × | 87.45 | 79.60 | 71.59 | 69.64 | 49.8 | 77.07 | G4 | √ | √ | √ | × | 86.54 | 86.89 | 69.94 | 76.82 | 52.9 | 80.05 | G5 | √ | √ | √ | √ | 87.46 | 87.34 | 76.00 | 80.20 | 52.9 | 82.75 |
|
Table 2. Comparison of ablation experiments of each improved module
Method | PL | SH | SV | LV | mAP |
---|
FR-O[23] | 79.42 | 37.16 | 35.30 | 38.02 | 47.48 | IE-Net[24] | 80.20 | 52.58 | 49.71 | 65.01 | 61.88 | SCRDet | 89.98 | 72.41 | 68.36 | 60.32 | 72.77 | RSDet[25] | 90.10 | 73.60 | 70.20 | 78.70 | 78.15 | R-YOLOv5m | 87.46 | 87.34 | 76.00 | 80.20 | 82.75 |
|
Table 3. Comparison of results of different algorithms on DOTA dataset subset
Method | R2CNN[26] | R2PN[27] | Gliding Vertex[28] | R-YOLOv5m |
---|
mAP /% | 73.07 | 79.60 | 88.20 | 88.89 |
|
Table 4. Comparison of different algorithms on HRSC2016 ship dataset