Author Affiliations
1Key Laboratory of Wireless Sensor Networks and Communication, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 201800, China2University of Chinese Academy of Sciences, Beijing 100864, China3Wuxi Hi-Tech Nano SensoringNet R&D Center of Chinese Academy of Sciences, Wuxi, Jiangsu 214135, Chinashow less
Fig. 1. Flow chart of proposed algorithm
Fig. 2. Improved fire detection model
Fig. 3. Feature extraction network
Fig. 4. Short connection branch network
Fig. 5. Dilated convolution module
Fig. 6. Attention mechanism network
Fig. 7. Bi-FPN
Fig. 8. Part of sample dataset. (a)-(c) Positive samples; (d)-(f) negative samples
Fig. 9. Partial test results
Parameter | ResNet-18 | ResNet-34 | ResNet-50 |
---|
MAE | 0.0101 | 0.0138 | 0.0104 | Fβ_fg | 0.8812 | 0.8815 | 0.8943 | Fβ_bg | 0.9010 | 0.8802 | 0.8734 |
|
Table 1. Performance analysis of feature extraction networks with different depth
Parameter | Proposed algorithm | YOLOv4 | RetinaNet |
---|
Precision | 0.890 | 0.900 | 0.802 | Recall | 0.880 | 0.880 | 0.773 | Fβ | 0.884 | 0.889 | 0.788 |
|
Table 2. Performance comparison between proposed algorithm and YOLOv4, RetinaNet
Parameter | Proposed algorithm | BASNet | PICANet |
---|
Fβ | 0.881 | 0.877 | 0.861 | MAE | 0.010 | 0.013 | 0.023 |
|
Table 3. Performance comparison between proposed algorithm and BASNet, PICANet
Parameter | Proposed algorithm | YOLOv4 | RetinaNet | BASNet | PICANet |
---|
Detection speed /(frame⋅s-1) | 4 | 4 | 4.5 | 2 | 3 | Model size /MB | 121.5 | 256 | 145.7 | 348.5 | 188.9 |
|
Table 4. Comparison of reasoning speed and model size of different algorithms
Parameter | Proposed algorithm | YOLOv4 |
---|
| 0.020 | 0.200 | | 0.003 | 0.067 | | 0.980 | 0.800 | | 0.997 | 0.933 |
|
Table 5. Comparison of detection results between proposed algorithm and YOLOv4
Single ResNet | Double | Bi-FPN | Dilated Conv | Attention mechanism | MAE | Fβ |
---|
| √ | | | | 0.085 | 0.8079 | | √ | √ | | | 0.043 | 0.8521 | | √ | √ | √ | | 0.032 | 0.8658 | | √ | √ | √ | √ | 0.010 | 0.8840 | √ | | √ | √ | √ | 0.032 | 0.8498 |
|
Table 6. Ablation experiments on flame dataset