Figure 1.
Flowchart of YOLOv5s algorithm
Figure 2.
Multi-scale detection structure
Figure 3.
Structure of CBAM module
Figure 4.
Channel Module
Figure 5.
Space module
Figure 6.
Schematic diagram of intersection and concatenation of IoU prediction and real frames.
Figure 7.
GIoU penalty content for minimising the area of the shaded region
Figure 8.
Computational schematic of MPDIoU
TABLE I.
Effect of internal parameters on the model
| Batch Size | Average accuracy | accuracy | recall rate | confidence level |
|---|---|---|---|---|
| Mean average precision | Precision P/% | Rcall | (math.) | |
| mAP percent | R/% | Confidence/% | ||
| 10 | 87.4 | 92.0 | 96.1 | 86.0 |
| 13 | 88.3 | 93.4 | 96.0 | 84.0 |
| 16 | 88.7 | 93.7 | 95.0 | 84.0 |
| 18 | 89.4 | 94.3 | 95.9 | 82.0 |
| 20 | 90.3 | 95.2 | 95.7 | 86.0 |
TABLE II.
Incorporation of multiple attention mechanisms
| Attention Mechanism | mAP% | P/% | R/% |
|---|---|---|---|
| +SE | 85.3 | 91.8 | 95.0 |
| +ECA | 86.6 | 92.3 | 94.1 |
| +CCA | 86.4 | 92.0 | 94.6 |
| +SA-Net | 87.8 | 92.5 | 94.7 |
| +MS-CAM | 87.5 | 92.7 | 95.2 |
| +CBAM | 88.5 | 93.2 | 95.8 |
TABLE III.
Comparison of different algorithms
| Model | Volume | mAP@0.5% |
|---|---|---|
| YOLOv5s | 14.0 | 86.9 |
| YOLOv8s | 22.4 | 86.5 |
| YOLOv6s | 37.4 | 83.0 |
| YOLOv4 | 245.9 | 79.5 |
| SSD | 100.3 | 71.0 |
| YOLOv51 | 93.7 | 89.7 |
| ours | 15.7 | 90.9 |
TABLE IV.
Ablation experiments
| Methods | X | C | L | Person/% | Car/% | mAP% |
|---|---|---|---|---|---|---|
| YOLOv5s | 79.2 | 94.3 | 86.9 | |||
| X-YOLOv5s | √ | 82.3 | 94.4 | 88.3 | ||
| C-YOLOv5s | √ | 81.8 | 95.2 | 88.5 | ||
| L-YOLOv5s | √ | 82.5 | 95.5 | 89.0 | ||
| XC-YOLOv5s | √ | √ | 83.1 | 96.2 | 89.6 | |
| XL-YOLOv5s | √ | √ | 83.4 | 96.6 | 90.0 | |
| CL-YOLOv5s | √ | √ | 83.9 | 96.3 | 90.1 | |
| XCL-YOLOv5s | √ | √ | √ | 84.7 | 97.0 | 90.9 |
Figure 9.
Comparison between before and after algorithm improvement