Author Affiliations
School of Information and Control Engineering, Xi'an University of Architecture and Technology, Xi'an, Shaanxi 710055, Chinashow less
Fig. 1. CNN structure diagram
Fig. 2. Overall framework of proposed algorithm
Fig. 3. Flow chart of proposed algorithm
Fig. 4. Flow chart of feature extraction process
Fig. 5. VGG16 network structure diagram
Fig. 6. ResNet50 structure diagram
Fig. 7. Partial sample datasets. (a)-(e) Positive samples; (f)-(i) negative samples
Fig. 8. Convolutional layers output feature maps. Shallow convolution layer: (a) network 1, (c) network 2; deep convolution layer: (b) network 1, (d) network 2
Fig. 9. Variation curve of flame detection performance with number of features
Method | Number of features | ACCR /% | DR /% | FAR /% |
---|
Method 1 | 117 | 95.00 | 90.0 | 0 | Ours | 16 | 98.25 | 96.5 | 0 |
|
Table 1. Influence of different feature selection algorithms on recognition
Method | Number offeatures | ACCR /% | DR /% | FAR /% |
---|
Network 1 | 32 | 93.50 | 92.0 | 5.0 | Network 2 | 27 | 96.00 | 95.5 | 3.5 | Ours | 16 | 98.25 | 96.5 | 0 |
|
Table 2. Influence of CNN serial fusion features on recognition
Algorithm | Number offeatures | ACCR /% | DR /% | FAR /% |
---|
Algorithm 1 | 3 | 76.25 | 84.0 | 31.5 | Algorithm 2 | 16 | 98.25 | 96.5 | 0 | Ours | 19 | 99.75 | 100.0 | 0.5 |
|
Table 3. Performance comparison of three detection algorithms
Algorithm | ACCR /% | DR /% | FAR /% |
---|
Ref. [7] | 86.88 | 91.85 | 18.10 | Ref. [9] | 96.73 | 94.78 | 1.32 | Ref. [14] | 97.28 | 98.25 | 3.70 | Ours | 99.75 | 100.00 | 0.50 |
|
Table 4. Results comparison of different detection algorithms