Author Affiliations
1China Airborne Missile Academy, Luoyang, Henan 471009, China2Aviation Key Laboratory of Science and Technology on Airborne Guided Weapons, Luoyang, Henan 471009, Chinashow less
Fig. 1. Proposed detection scheme
Fig. 2. Detection schemes of different algorithms
Fig. 3. Typical fully convolutional neural networks for infrared small target detection.
Fig. 4. Proposed fully convolutional neural network
Fig. 5. Typical target image based on 2D Gaussian model
Fig. 6. Typical infrared images with small targets
Fig. 7. Segmentation results of different networks
Fig. 8. ROC curve of proposed algorithms with and without contrast feature
Fig. 9. Test images and results of seven algorithms
Fig. 10. ROC curves of different algorithms
Type | Description | Pd | Fa | GFLOPs | M1 | Base network in Fig. 4 | 0.940 8 | 7.97×10-6 | 0.235 | M11 | Without down-sampling layer | 0.938 9 | 2.01×10-5 | 0.494 | M12 | Without feature of 3rd conv layer | 0.929 4 | 1.11×10-5 | 0.226 | M13 | Without SE layer | 0.938 9 | 1.42×10-5 | 0.235 |
|
Table 1. Comparation of different design choices of proposed network
Type | Description | Pd | Fa | GFLOPs | M1 | Network in Fig. 4 | 0.940 8 | 7.97×10-6 | 0.235 | M2 | Network in Fig. 3(a) | 0.931 3 | 1.06×10-5 | 6.47 | M3 | Network in Fig. 3(b) | 0.933 2 | 4.23×10-5 | 1.54 |
|
Table 2. Performance of different networks
Index | ADMD | IPIM | LIG | MaxMedian | FMDNN | FCN+CNN | S-FCN | 1 | 20.3 | 78.9 | 31.0 | 12.0 | 25.1 | 63.6 | 118.0 | 2 | 17.9 | 0 | 0.1 | 4.8 | 15.2 | 21.0 | 28.9 | 3 | 10.1 | 38.9 | 6.4 | 12.7 | 5.3 | 17.1 | 55.9 | 4 | 16.9 | 17.9 | 20.5 | 15.9 | 19.0 | 25.9 | 25.3 | 5 | 8.2 | 12.5 | 9.6 | 1.8 | 5.9 | 6.9 | 12.6 | 6 | 56.1 | 0.0 | 1.9 | 50.3 | 33.5 | 310.8 | 520.7 | 7 | 5.2 | 24.2 | 3.8 | 19.5 | 7.2 | 9.9 | 29.8 | 8 | 12.5 | 37.0 | 35.2 | 14.5 | 4.4 | 19.7 | 51.3 |
|
Table 3. G of different algorithms on 8 test images
Index | ADMD | IPIM | LIG | MaxMedian | FMDNN | FCN+CNN | S-FCN | 1 | 4.2 | 19.9 | 6.3 | 2.5 | 5.1 | 12.8 | 23.7 | 2 | 5.3 | - | - | 1.4 | 4.5 | 6.2 | 8.5 | 3 | 4.1 | 19.6 | 5.0 | 5.1 | 2.1 | 6.8 | 22.3 | 4 | 6.1 | 8.2 | 7.4 | 5.8 | 6.9 | 9.4 | 9.2 | 5 | 2.2 | 3.4 | 2.6 | 0.9 | 1.6 | 1.9 | 3.4 | 6 | 3.7 | - | - | 3.0 | 1.9 | 17.9 | 29.3 | 7 | 2.3 | 13.6 | 3.1 | 8.3 | 3.1 | 4.3 | 12.7 | 8 | 6.6 | 27.7 | 18.4 | 7.6 | 2.4 | 10.3 | 26.8 |
|
Table 4. B of different algorithms on 8 test images
Method | ADMD | IPIM | LIG | MaxMedian | FMDNN | FCN+CNN | S-FCN | Time/s | 0.996 | 0.866 | 0.284 | 0.219 | 0.066 | 0.164 | 0.033 |
|
Table 5. Average run time of different algorithms