Figure 1.
Part of the image data used for training: (a) Fire with poor resolution. (b) Fire in a small area. (c) Fire with flame obstruction. (d) Fire disturbed by smoke.
Figure 2.
YOLOv 5 input data enhancement method
Figure 3.
CBAM overall structure
Figure 4.
CAM structure
Figure 5.
SAM structure
Figure 6.
Ordinary convolution
Figure 7.
Depth separation convolution: (a) Depth convolution. (b) Pointwise convolution
Figure 8.
Improved model framework
Figure 9.
The position of CBAM in YOLOv5s 6.0 version
TABLE I.
DATASET SETTINGS
| Dataset | Training | Test | Validation | Total |
|---|---|---|---|---|
| Homemade forest fire data set | 1442 | 617 | 617 | 2676 |
| Other institutes data set | 600 | 200 | 200 | 1000 |
TABLE III.
TRAINING SETTINGS PARAMETERS
| Training parameters | Detail |
|---|---|
| Epochs | 100 |
| Batch-size | 16 |
| Image-size | 6 40 × 640 |
| Initial learning rate | 0.01 |
| Optimization algorithm | SGD |
TABLE II.
EXPERIMENTAL SETTINGS
| Lab Environment | Detail |
|---|---|
| programming language | Python3.8.5 |
| operating system | Windows 10 |
| deep learning framework | Pytorch 1.8.0 |
| GPU | 4x NVIDIA TITIAN V |
TABLE IV.
COMPARATIVE TEST RESULTS OF THE MODEL
| Model | P | R | FPS |
|---|---|---|---|
| YOLOv5s | 0.811 | 0.786 | 59 |
| YOLOv5s + CBAM | 0.814 | 0.790 | 60 |
| YOLOv5s + SE | 0.810 | 0.787 | 5 8 |
| YOLOv5s +ECA | 0.812 | 0.791 | 5 9 |
| YOLOv5s + dsCBAM | 0.812 | 0.787 | 62 |
| YOLOv5s + dsCBAM +Alpha-IoU | 0.821 | 0.813 | 61 |
| YOLOv5s + dsCBAM + SIoU | 0.860 | 0.834 | 60 |
| YOLOv5s + dsCBAM+ VariFocal (Ours) | 0.871 | 0.816 | 64 |
Figure 10.
Experimental identification results: (a) Improved model. (b) Original model.
Figure 11.
Experimental Experimental results of misdetection of forest street lights at night. (a) Original model. (b) Improved model.