Figure 1.
RT-DETR structure diagram
Figure 2.
Improved RT-DETR algorithm
Figure 3.
SCSA overall structure diagram. (a) Shows the structure diagram of SCSA, and (b) shows the structure diagram of MS-DWConv1d
Figure 4.
CA-SHSA Module Structure
Figure 5.
CARAFE Module Structure Diagram
Figure 6.
Structural diagram of RepC3
Figure 7.
BFAM Structure Diagram
Figure 8.
MFRC3 Module Structure Diagram
Figure 9.
Target category and size distribution chart
TABLE I.
EXPERIMENTAL HYPERPARAMETERS
| Hyperparameters | Configurations |
|---|---|
| batch_size | 4 |
| epoch | 200 |
| optimizer | AdamW |
| num_workers | 4 |
| image_size | 640×640 |
TABLE II.
The effects of each improvement module
| BasicSCSABlock | CARAFE | MFRC3 | Param/MB | APs/% | APm/% | APl/% | mAP@0.5/% | mAP@0.5:0.95/% | FPS |
|---|---|---|---|---|---|---|---|---|---|
| - | - | - | 19.2 | 19.2 | 39.3 | 54.8 | 48.0 | 28.5 | 81.4 |
| √ | - | - | 18.1 | 19.3 | 39.4 | 54.8 | 48.1 | 28.6 | 80.5 |
| - | √ | - | 18.2 | 19.3 | 39.5 | 57.1 | 48.5 | 28.9 | 70.8 |
| - | - | √ | 34.2 | 20.3 | 42.0 | 56.6 | 50.5 | 30.4 | 49.5 |
| √ | √ | - | 18.2 | 18.9 | 38.9 | 53.2 | 47.6 | 28.3 | 74.7 |
| - | √ | √ | 34.3 | 20.5 | 41.7 | 53.3 | 51.0 | 30.1 | 45.6 |
| √ | - | √ | 27.9 | 21.0 | 42.3 | 52.8 | 50.0 | 29.5 | 63.2 |
| √ | √ | √ | 28.0 | 21.0 | 42.4 | 58.3 | 51.1 | 31.0 | 53.7 |
Figure 10.
mAP curve comparison chart
Figure 11.
Visualization of experimental results comparison
TABLE III.
Comparison with mainstream algorithms
| model | Param/MB | GFLOPs | FPS | mAP@0.5% | mAP@0.5 : 0.95% |
|---|---|---|---|---|---|
| YOLO v5x | 86.2 | 203.8 | 34 | 42.5 | 25.2 |
| YOLO v8n | 3.0 | 8.1 | 119 | 33.5 | 17.8 |
| YOLO v8s | 11.2 | 28.7 | 85 | 37.3 | 21.6 |
| YOLO v10n | 2.69 | 8.2 | 111 | 29.9 | 17.1 |
| YOLO v10s | 8.04 | 24.5 | 143 | 36.4 | 21.4 |
| Defor mable | 40 | 196 | 29 | 42.2 | 27.1 |
| DETR DINO | 47 | 279 | 24 | 46.2 | 29.4 |
| RT-DETR | 19.2 | 30.2 | 81.4 | 48.0 | 28.5 |
| Ours | 28.0 | 56.0 | 53.7 | 51.1 | 31.0 |