Research Article
Predicted Anchor Region Proposal with Balanced Feature Pyramid for License Plate Detection in Traffic Scene Images
Table 5
Detection results of each proposed network on CCPD dataset.
| Network | Detection performance (%) | Base | DB | FN | Rotate | Tilt | Weather | Challenge | Average |
| Faster R-CNN | 98.1 | 92.1 | 83.7 | 91.8 | 89.4 | 81.1 | 83.9 | 88.6 | FPN with faster R-CNN baseline | 99.2 | 95.4 | 87.5 | 93.0 | 91.3 | 85.4 | 86.3 | 91.2 | Faster R-CNN + balanced feature pyramid | 99.5 | 96.2 | 89.1 | 93.2 | 91.6 | 88.9 | 90.1 | 92.7 | Faster R-CNN + balanced pyramid without L2-norm | 99.3 | 96.2 | 88.9 | 93.1 | 91.0 | 87.6 | 88.3 | 92.1 | Faster R-CNN + predicted anchor RPN | 98.5 | 92.6 | 84.0 | 91.5 | 89.4 | 81.4 | 84.4 | 88.8 | Faster R-CNN + balanced pyramid with L2-norm + predicted anchor RPN | 99.5 | 96.4 | 90.1 | 93.2 | 91.8 | 89.7 | 91.2 | 93.1 |
|
|