Interferometric Wavefront Sensing System Based on Deep Learning

Round 1

Reviewer 1 Report

Few questions and comments are as following:

 

1. (Line 77 in p.2) Lens 1, 2 of are not shown in phase-shifting interferometer of Fig. 1.
2. (Line 154 in p.4) Authors shall clarify if there is any clue or physical attribute that the phase wrap count was limited in [-5, 5] in the training set? Since according to the experiment shown in Fig. 9 that the testing wavefront was merely constrained in [0, pi].
3. (Line 156 in p.4) What does “smooth method” mean in the Net1 output?     
4. (Line 185 in p.5) Authors shall clarify how the Covariance Matrix Method was validated to fit the first 36 Zernike coefs with true phase map by 2D Gaussian generation.
5. (Line 189 in p.6) Any parameters for the salt and pepper noise?
6. (Eq. 9 in p.6) same as Eq. (8).
7. (Fig. 4 in p.7) There appears some spikes in Difference. Where does the spike come from? As decreasing SNR in Fig. 5, more spikes in difference would severely influence the estimation accuracy of Zernike coef. Authors shall explain the denoising features about Net2.
8. (Fig. 6 in p.8) In the Zernike coef fitting results, only two wavefronts were tested. How to justify your assumption that Net2 outperforms the least square. Especially in smooth wavefront (fewer Zernike coef), the performance of least square might be comparable to network, to my knowledge.  
9. (Fig. 9 in p.10) (e) authors shall reveals all the Zernike coefs. It seems that most high-order Zernike coefs are insignificant. (f) the experimentally distorted wavefront scales in [0, pi], much less than the simulated training dataset (Fig. 3d). Is it possible to conduct a wavefront distortion with more-than-one wrap count? 

 

 

Author Response

We appreciate you deeply for your comments concerning our manuscript entitled Interferometric wavefront sensing system based on deep learning (Manuscript ID: applsci-982522). The paper has been revised according to your suggestions one by one.

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

The authors present a wavefront detection system that utilizes deep learning. They design a rather simple experimental setup consisting of four discrete reference phase shifts and then employ two neural networks. The first network performs phase wrapping, while the second network takes the phase map as an input and provides the Zernike coefficient as an output. When testing the neural network on simulation data, it performs well. When using real experimental data, however, the RMSE of the reconstructed wavefront degrades to 0.61, which is okay, but not overly convincing.

I do not think that the state of the art in the field is necessarily represented well in the current manuscript. For example robustness to noise when using neural networks has been discussed and demonstrated in detail, see e.g. Paine and Fienup, Optics Letters 43, 1235 (2018). Further, in terms of robustness, I would assume that the interferometric approach employed here is a drawback compared to intensity-based techniques. The authors might want to put forward some specific application scenarios, where their approach may prove competitive. Contrary to the authors' claim, other approaches are not way more sensitive to external noise.

Overall, the results seem correct to me and the manuscript is reasonably well written, but the manuscript is somewhat lacking as the results presented here are either not too convincing or not really motivated well. A revised version that removes some errors and problems with the figures and explicitly mentions some realistic application scenario might be publishable in Appl. Sci., though.

Some minor comments:
- In figure 1, it is almost impossible to read the red letters that overlap with the darker region of the optical table, e.g. M3. I suggest to use a different color or to place the letters elsewhere.

- In figure 1 and figure 7 a mirror M1 is mentioned in the text, but never marked in the figure. In figure 7, "PLS" is shown in the figure, but the main text never mentions that "PLS" is short for precision linear stage.

- Several of the figures are too small to recognize any detail, e.g. figures 4 and 5.

- For real applications, the most important figure of merit will be, how quickly wavefronts may be corrected. Can the authors comment on that? Most likely, the need to actively shift the reference phase will render this approach rather slow.

- In the discussion of figure 9, the possible reasons for the quite substantial noise observed at the borders of the beam are not discussed. Maybe it would be helpful to do so.

Author Response

We appreciate you deeply for your comments concerning our manuscript entitled Interferometric wavefront sensing system based on deep learning (Manuscript ID: applsci-982522). The paper has been revised according to your suggestions one by one.

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

With respect to the part of the manuscript that focuses on optics, it is my opinion that the manuscript is okay now. With respect to the other topics the manuscript touches (machine learning, algorithms, ...), I do not feel competent enough to decide whether the manuscript has improved and would leave the decision to the other referee.