Author Affiliations
1Shanghai Institute of Measurement and Testing Technology, Shanghai 201203, China2Shanghai Key Laboratory of Online Testing and Control Technology, Shanghai 201203, China3School of Computer Science and Information Security, Guilin University of Electronic Technology, Guilin 541004, Chinashow less
Fig. 1. (a) Absolute phase (real data set); (b) Wrapped phase (real data set); (c) Absolute phase (simulated data set 1, 2); (d) Wrapped phase (simulated data set 1); (e) Wrapped phase (simulated data set 2); (f) Wrapped phase (simulated data set 3)
Fig. 2. Comparison of Shannon entropy between real and simulated dataset
Fig. 3. Network model diagram
Fig. 4. Schematic diagram of CBiLSTM module
Fig. 5. Attention gate schematic
Fig. 6. The performance of each network model under different noise conditions
Fig. 7. The performance of each network model in the case of discontinuity
Fig. 8. Performance of individual network models in case of aliasing
Fig. 9. The generalization ability of each network model in the real data set
Fig. 10. Different performance comparison of each network model
Fig. 11. Comparison between the predicted phase and the true absolute phase of the network model proposed in this article in the above three situations
Fig. 12. Experimental test system
Fig. 13. Data acquisition diagram
Fig. 14. Horizontal and vertical phase unwrapping on real data sets
Fig. 15. Picture of the experimental situation with dust and scratches on the lens
Fig. 16. Phase unwrapping on complex real datasets
Network model | Noise NRMSE | Discontinuous NRMSE | Aliased NRMSE | Our Net(Loss=Lmeaw) | 2.05% | 2.93% | 2.64% | Our Net(Loss=Lerror) | 1.49% | 2.99% | 2.35% | U-Net[20] | 14.03% | 12.59% | 13.48% | Wang et al[13] | 13.27% | 11.6% | 12.24% | Res-UNet[21] | 13.26% | 14.69% | 12.98% | Our Net | 1.12% | 1.81% | 1.68% | Perera et al[19] | 1.46% | 2.09% | 1.87% |
|
Table 1. Performance comparison of various network models in terms of errors
Network model | Noise | | Discontinuous | | Aliased | Time/s | Total time/s | | Time/s | Total time/s | | Time/s | Total time/s | U-Net[20] | 7 | 2625 | | 6 | 2406 | | 6 | 2124 | Wang et al[13] | 11 | 3256 | | 12 | 3996 | | 11 | 3080 | Res-UNet[21] | 16 | 6384 | | 21 | 6153 | | 17 | 6443 | Our Net | 15 | 1275 | | 18 | 2250 | | 13 | 2860 | Perera et al[19] | 18 | 1746 | | 18 | 2574 | | 17 | 4709 |
|
Table 2. Performance comparison of various network models in terms of time consumption
Serial number | Based on improved U-Net | Model CBiLSTM | Soft attention | NRMSE | PSNR | SSIM | 1 | √ | | | 10.06% | 13.8 | 0.46 | 2 | √ | √ | | 0.92% | 35.78 | 0.96 | 3 | √ | | √ | 1.16% | 34.26 | 0.94 | 4 | √ | √ | √ | 0.75% | 40.87 | 0.98 |
|
Table 3. Quantitative comparison of ablation experiments on noisy datasets