Author Affiliations
1School of Electrical and Control Engineering, Xi'an University of Science and Technology, Xi'an, Shaanxi 710054, China2School of Safety Science and Engineering, Xi'an University of Science and Technology, Xi'an, Shaanxi 710054, Chinashow less
Fig. 1. Structure of Inception module
Fig. 2. Structure of InceptionV1 module
Fig. 3. Structure of improved Inception module
Fig. 4. Front-end structure of improved InceptionV1 module
Fig. 5. Overall structure of improved InceptionV1 module
Fig. 6. Feature extraction framework
Fig. 7. Results of model parameter complexity optimization experiment
Fig. 8. Final network structure
Fig. 9. Detection results of flame super-pixel. (a) Original image; (b) video super-pixel segmentation result; (c) full-frame detection result; (d) final detection result of removing non-flame area
Fig. 10. Ground truth annotation image
Fig. 11. Partial flame detection results of proposed method. (a) Image 1; (b) image 2; (c) image 3; (d) image 4; (e) image 5; (f) image 6
Dataset | Quantity |
---|
ImageNet dataset | 3067 | Bilkent University Fire dataset | 4782 | Durham University Fire dataset | 4563 | Non-fire dataset | 5439 |
|
Table 1. Flame image data source and quantity
Imp-A | Imp-B | Focal-Loss | FLOPS/109 | Test accuracy /% |
---|
| | | 1.502 | 94.82 | | | √ | 1.502 | 95.09 | √ | | √ | 1.232 | 96.23 | √ | | | 1.232 | 95.87 | | √ | | 1.467 | 95.24 | | √ | √ | 1.467 | 96.12 | √ | √ | | 1.197 | 96.56 | √ | √ | √ | 1.197 | 97.01 |
|
Table 2. Results of ablation experiments
Method | TPR | FPR | A | P | F1-score |
---|
AlexNet | 0.91 | 0.19 | 0.88 | 0.92 | 0.91 | VGG-16[7] | 0.92 | 0.12 | 0.90 | 0.93 | 0.92 | InceptionV1 | 0.94 | 0.09 | 0.95 | 0.95 | 0.94 | Proposed method | 0.96 | 0.08 | 0.95 | 0.95 | 0.95 |
|
Table 3. Comparison of index evaluation of different methods
Method | C /106 | A /% | A∶C | FPS |
---|
AlexNet | 72.0 | 91.7 | 1.27 | 20.1 | VGG-16 | 215.3 | 92.6 | 1.36 | 13.8 | InceptionV1 | 6.1 | 95.2 | 15.60 | 40.7 | Ref. [20] | \ | \ | \ | 14.2 | Ref. [19] | \ | \ | \ | 3.9 | Proposed method | 1.8 | 95.4 | 53.00 | 108.4 |
|
Table 4. Evaluation and comparison of network computing performance
Method | TPR | FPR | F1-score | P | A |
---|
SEEDS+AlexNet | 0.86 | 0.29 | 0.84 | 0.83 | 0.86 | SEEDS+VGG16 | 0.85 | 0.31 | 0.82 | 0.80 | 0.85 | SEEDS+InceptionV1 | 0.94 | 0.18 | 0.92 | 0.91 | 0.94 | Ref. [20] | 0.91 | 0.13 | 0.88 | 0.86 | 0.87 | Ref. [19] | 0.87 | 0.17 | 0.88 | 0.90 | 0.90 | Proposed method | 0.94 | 0.04 | 0.94 | 0.95 | 0.96 |
|
Table 5. Performance evaluation results of flame detection model
Method | TPR | F1-score | P | S |
---|
SEEDS+AlexNet | 0.80 | 0.78 | 0.76 | 0.78 | SEEDS+VGG16 | 0.85 | 0.84 | 0.84 | 0.76 | SEEDS+InceptionV1 | 0.90 | 0.88 | 0.87 | 0.87 | Proposed method | 0.91 | 0.90 | 0.89 | 0.90 |
|
Table 6. Evaluation results of flame localization performance