Table 1.
Reference Training Configurations for Stage 1
Training Configuration | Reference Value |
---|
Number of samples | 10,000 |
Optimizer | AdamW [33] |
Cost function | Mean squared error |
Base learning rate | |
Weight decay | 0.1 |
Learning rate schedule | Cosine decay [34] |
Batch size | 128 |
Training epochs | 150 |
Table 2.
Reference Training Configurations for Stage 2
Training Configuration | Reference Value |
---|
Number of samples | 100,000 |
Optimizer | AdamW [33] |
Cost function | Mean squared error |
Base learning rate | |
Weight decay | 0.1 |
Learning rate schedule | Cosine decay [34] |
Batch size | 1024 |
Training epochs | 400 |
Table 3.
Performance Comparison of Different Networks
Networks | Center RMSE | Radius Relative MAE | Training Time | Demodulation Time (using CPU) | Demodulation Time (using GPU) |
---|
VGG with single head [30] | 51.06 pixels | 82.47% | 12.98 h | 82.42 ms | 4.98 ms |
VGG with multiple heads | 46.58 pixels | 16.70% | 11.55 h | 81.15 ms | 3.63 ms |
Residual network with single head | 66.27 pixels | 112.15% | 11.03 h | 39.22 ms | 5.88 ms |
Residual network with multiple heads (current work) | 0.23 pixels | 0.15% | 10.12 h | 38.87 ms | 6.67 ms |
Table 4.
Estimations of the Proposed Network in Fig. 7
Example | Image Scale (/pixel) | Center Ground Truth | Radius Ground Truth (m) | Center Estimation | Radius Estimation without MLP (m) | Radius Estimation with MLP (m) |
---|
Fig. 7(a) | 15.171 | (20, 107) | 1.6698 | (20, 107) | 1.8348 | 1.6754 |
Fig. 7(b) | 16.667 | (20, 107) | 1.6698 | (20, 107) | 1.6649 | 1.6649 |
Fig. 7(c) | 19.541 | (20, 107) | 1.6698 | (20, 107) | 1.4217 | 1.6724 |
Table 5.
Estimations of the Proposed Network in Fig. 8
Example | Image Scale (/pixel) | Center Ground Truth | Radius Ground Truth (m) | Center Estimation | Radius Relative MAE |
---|
Fig. 8(a) | 15.685 | (82, 106) | 0.855 | (81, 105) | 2.89% |
Fig. 8(b) | 21.053 | (110, 94) | 1.443 | (111, 94) | 1.59% |
Fig. 8(c) | 15.685 | (73, 118) | 1.443 | (73, 118) | 0.37% |
Table 6.
Estimation of the Proposed Network in Fig. 9
Image Scale (/pixel) | Center Ground Truth | Radius Ground Truth (mm) | Center Estimation | Radius Relative MAE |
---|
1.3612 | (121, 117) | 10.02 | (121, 116) | 0.45% |
Table 7.
Estimations of the Proposed Network in Fig. 11
Example | Image Scale (/pixel) | Center Ground Truth | Radius Ground Truth (m) | Center Estimation | Radius Relative MAE |
---|
Fig. 11(a) | 15.685 | (73, 118) | 1.443 | (73, 118) | 0.37% |
Fig. 11(b) | 17.538 | (138, 139) | 1.443 | (139, 137) | 2.23% |
Fig. 11(c) | 17.538 | (118, 131) | 1.443 | (117, 133) | 4.51% |
Table 8.
Noise Robustness Comparison
Setups | Center RMSE | Radius Relative MAE |
---|
Trained without noise | 11.34 pixels | 8.28% |
Trained with noise | 0.34 pixels | 0.24% |
Table 9.
Estimations of the Proposed Network in Fig. 13
Example | Center Ground Truth | Radius Ground Truth (m) | Center Estimation | Radius Relative MAE |
---|
Fig. 13(a) | (114, 206) | 1.6092 | (114, 205) | 0.04% |
Fig. 13(b) | (173, 189) | 1.1145 | (173, 188) | 0.69% |
Fig. 13(c) | (70, 62) | 1.5889 | (70, 62) | 0.11% |
Table 10.
Estimations of the Proposed Network on Simulated Newton’s Ring Patterns with Different Gaussian Noise Intensities
SNR(dB) | Center Estimation | Radius Relative MAE |
---|
| (114, 206) | 0.09% |
30 | (114, 206) | 0.15% |
25 | (114, 206) | 0.20% |
20 | (114, 206) | 0.11% |
15 | (114, 206) | 0.07% |
10 | (114, 206) | 0.14% |
5 | (114, 206) | 0.10% |
0 | (114, 206) | 0.04% |
-5 | (114, 205) | 0.04% |
Table 11.
Estimations of the Proposed Network on Simulated Newton’s Ring Patterns with Different Speckle Noise Intensities
SNR(dB) | Center Estimation | Radius Relative MAE |
---|
| (70, 62) | 0.09% |
30 | (70, 62) | 0.24% |
25 | (71, 62) | 0.32% |
20 | (70, 62) | 0.32% |
15 | (70, 62) | 0.20% |
10 | (70, 62) | 0.29% |
5 | (70, 62) | 0.38% |
0 | (71, 62) | 0.51% |
-5 | (70, 62) | 0.11% |