Fig. 1. The proposed network architecture.
Fig. 2. The sketches of 3D covolution and factorized 3D convolution: (a) 3D covolution; (b) factorized 3D convolution.
Fig. 3. The example of different error distributions: (a) the ground truth; (b) 1th predicted result with uniform error; (c) 2th predicted result with concentrated error.
Fig. 4. The function of intensity weighting parameter.
Fig. 5. The example of samples: (a) the target samples; (b) the background samples.
Fig. 6. The original image and results of different methods for 1th Background.: (a) the original input; (b) the result of our method; (c) the result of Lin’s method; (d) the result of Max-Mean; (e) the result of TopHat; (f) the result of STDA.
Fig. 7. The original image and results of different methods for 2th Background.: (a) the original input; (b) the result of our method; (c) the result of Lin’s method; (d) the result of Max-Mean; (e) the result of TopHat; (f) the result of STDA.
Fig. 8. The original image and results of different methods for Target 1: (a) the original input; (b) the result of our method; (c) the result of Lin’s method; (d) the result of Max-Mean; (e) the result of TopHat; (f) the result of STDA.
Fig. 9. The original image and results of different methods for Target 2: (a) the original input; (b) the result of our method; (c) the result of Lin’s method; (d) the result of Max-Mean; (e) the result of TopHat; (f) the result of STDA.
Fig. 10. The display of Target 2: (a) the original gray image; (b) the standard deviation in the time domain.
Fig. 11. The ROC curves of different methods.
Fig. 12. The result of different input size: (a) the input image with 35×35 pixels; (b) the result of image with35×35 pixels; (c) the input image with 45×45 pixels; (b) the result of image with45×45 pixels.
Fig. 13. The ROC curves with different jitters.
Fig. 14. The ROC curves with different mean original SCRs.
Item | Flops(G) | Parameters(K) |
---|
3D Conv. | 0.06 | 28.67 | Factorized 3D Conv. | 0.03 | 9.38 |
|
Table 1. Computation comparison of different convolutions.
Item | Parameter |
---|
CPU | Intel i7,2.8GHz×12 | GPU | Nvidia-1080Ti | RAM | 64GB | System | Ubuntu 18.04 | Disk | 2TB | Software | Pytorch 1.1 | Language | Python 3.6 |
|
Table 2. The simulation environment.
| Proposed | Lin’s | Max-Mean | TopHat | STDA |
---|
Target 1 | 20.4061 | 16.8253 | 20.9308 | 20.0394 | 19..8378 | Target 2 | 19.0953 | 11.9228 | 3.4410 | 3.0623 | 16.5872 | Mean | 19.7507 | 14.3741 | 12.1859 | 11.5509 | 18.2125 |
|
Table 3. Background suppression comparison by SCR in output.
| Proposed | Lin’s | Max-Mean | TopHat | STAD |
---|
Target 1 | 1.3527 | 1.0293 | 1.2477 | 1.1210 | 1.3375 | Target 2 | 7.1815 | 4.3218 | 1.2520 | 1.0431 | 5.5936 | Mean | 4.2671 | 2.6755 | 1.2498 | 1.0821 | 3.4656 |
|
Table 4. Background suppression comparison by BSF.
| Proposed | Lin’s | Max-Mean | TopHat | STDA |
---|
Average runtime(s)/sample | 5.84×10-4 | 3.38×10-4 | 3.30×10-3 | 1.33×10-2 | 4.51×10-3 |
|
Table 5. Average runtime comparison.