Fig. 1. Detection scheme of proposed algorithm
Fig. 2. Detection scheme of typical FCN
Fig. 3. Detection scheme of MDvsFA_cGAN
Fig. 4. Network structure of proposed algorithm
Fig. 5. Overall flow of CBAM
Fig. 6. Overall flow of CAM
Fig. 7. Overall flow of SAM
Fig. 8. Real infrared images and binarised labels
Fig. 9. Point source generated by PSF combined with rotation transformation
Fig. 10. Synthetic infrared images and labels based on PSF
Fig. 11. Synthetic infrared images and labels based on Mosaic
Fig. 12. Correspondence of indicator variable
Fig. 13. Effects comparison of different epoch
Fig. 14. RF_measure index comparison of typical deep learning model. (a) RF_measure metric iterations for each deep learning model on the IR_GS Dataset; (b) RF_measure metric iterations for each deep learning model on the IR_GMS Dataset
Fig. 15. Comparison of detection results for different algorithms
Fig. 16. Given infrared images and predicted results
Combination | Layer | Kernel_size | Stride | Padding | Dilation | In_channel | Out_channel | ReLU_param |
---|
Encoder | CBL_1 | 3 | 1 | 1 | 1 | 3 | 64 | 0.2 | CBL_2 | 3 | 1 | 2 | 2 | 64 | 64 | 0.2 | CBL_3 | 3 | 1 | 4 | 4 | 64 | 64 | 0.2 | CBL_4 | 3 | 1 | 8 | 8 | 64 | 64 | 0.2 | CBL_5 | 3 | 1 | 16 | 16 | 64 | 64 | 0.2 | CBL_6 | 3 | 1 | 32 | 32 | 64 | 64 | 0.2 | CBL_7 | 3 | 1 | 64 | 64 | 64 | 64 | 0.2 | Decoder | CBL_8 | 3 | 1 | 32 | 32 | 64 | 64 | 0.2 | CBL_9 | 3 | 1 | 16 | 16 | 128 | 64 | 0.2 | CBL_10 | 3 | 1 | 8 | 8 | 128 | 64 | 0.2 | CBL_11 | 3 | 1 | 4 | 4 | 128 | 64 | 0.2 | CBL_12 | 3 | 1 | 2 | 2 | 128 | 64 | 0.2 | CBL_13 | 3 | 1 | 1 | 1 | 128 | 64 | 0.2 |
|
Table 1. Design parameters of encoding-decoding
Dataset | Synthesis | Mosaic augmentation | Total | Training set | Valid set | LSNR /dB |
---|
IR_GS Dataset | √ | | 10000 | 9900 | 100 | 2.83 | IR_GMS Dataset | √ | √ | 15000 | 14950 | 150 | 2.96 |
|
Table 2. Dataset expansion mode and details
Dataset | Network | RFPR | Rprecision | Rrecall | RF_measure | Speedup |
---|
IR_GS Dataset | MDvsFA_cGAN | 1.48×10-3 | 0.740 | 0.380 | 0.502 | 1.6X | DeepLab_ResNet | 7.32×10-4 | 0.240 | 0.189 | 0.212 | 5.1X | Fcn8x_ResNet | 4.23×10-4 | 0.380 | 0.282 | 0.324 | 2.8X | SegNet | 4.36×10-4 | 0.581 | 0.350 | 0.436 | 1.6X | UNet | 3.21×10-4 | 0.526 | 0.390 | 0.448 | 1.2X | UNet++ | 3.16×10-4 | 0.624 | 0.414 | 0.498 | 3.5X | dilation | 5.64×10-4 | 0.723 | 0.403 | 0.518 | 1X | dilation_CBAM | 4.53×10-4 | 0.747 | 0.445 | 0.558 | 1.2X | IR_GMS Dataset | MDvsFA_cGAN | 1.04×10-3 | 0.683 | 0.500 | 0.578 | 1.6X | DeepLab_ResNet | 4.44×10-4 | 0.330 | 0.342 | 0.336 | 5.1X | Fcn8x_ResNet | 5.01×10-4 | 0.452 | 0.286 | 0.350 | 2.8X | SegNet | 3.49×10-4 | 0.607 | 0.489 | 0.542 | 1.6X | UNet | 3.86×10-4 | 0.597 | 0.465 | 0.522 | 1.2X | UNet++ | 3.17×10-4 | 0.584 | 0.483 | 0.529 | 3.5X | dilation | 5.46×10-4 | 0.693 | 0.542 | 0.608 | 1X | dilation_CBAM | 4.71×10-4 | 0.670 | 0.588 | 0.626 | 1.2X |
|
Table 3. Comparison of experimental results of typical deep learning model