Figure 1.
Road surface damage dataset under different conditions
Figure 2.
RT-DETR-r18 model structure
Figure 3.
Improved RT-DETR model structure
Figure 4.
Diagrams of different types of cracks
Figure 5.
Structure of LMBA module
Figure 7.
Structure of EFKM module
Figure 6.
Comparison of feature Pyramid net
TABLE I.
EXPERIMENTAL ENVIRONMENT
| Experimental environment | Version |
|---|---|
| CPU | Intel Xeon Platinum 8352V |
| GPU | NVIDIA GeForce RTX4090D |
| Language | Python3.9 |
| Deep Learning Framework | Pytorch1.13.1 |
| CUDA | 11.6.0 |
TABLE II.
DISEASE CATEGORY
| Category | Train Set | Test Set |
|---|---|---|
| D00(Longitudinal cracks) | 7419 | 876 |
| D10(Transverse cracks) | 5702 | 636 |
| D20(Alligator cracks) | 6244 | 689 |
| D40(Potholes) | 2316 | 248 |
TABLE III.
COMPARISON BEFORE AND AFTER IMPROVEMENT
| Algorithm | Pars/M | FLOPS/G | FPS/f/s | mAP/% |
|---|---|---|---|---|
| RT-DETR | 19.8 | 57.3 | 69 | 67.1 |
| Improved RT-DETR | 14.6 | 45.2 | 60 | 69.2 |
Figure 8.
Comparison chart of mAP during training
Figure 9.
Average precision of each label in RT-DETR
Figure 10.
Average precision of each label in Improved RT-DETR
Figure 11.
Visual comparison of test results