Denoising of Wrapped Phase in Digital Speckle Shearography Based on Convolutional Neural Network

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The study introduces a CNN model called SDCNN, aiming to balance speckle noise reduction and boundary preservation in defect detection using speckle shear interferometry. Using techniques like cosine regularization, a more flexible speckle noise simulation approach, and a specific loss function, SDCNN shows outstanding performance in four-step phase-shifting simulated images and real-world speckle noise scenarios, providing valuable insights for future research.

It looks like the results are good. I would be interested in knowing more about the various techniques the authors used to get their findings.

Comments on the Quality of English Language

Minor editing of English language required.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes a trained approach to unwrap noisy phase maps obtained from shearing imaging system. The key idea is to train a denoiser on simulated data that takes the noisy and wrapped phase map as input, and outputs a noise-free wrapped phase map. This can then be fed into a secondary phase unwrapping algorithm that produces the final result. Experimental results show that the proposed approach is as good or better than state-of-art denoiser, DNCNN.

 

While the core idea is promising, and the problem is well motivated, the paper is simply missing too many details. To an outsider, even the optical setup is unclear, although a couple of equations that resemble a Michelson interferometer have been provided. There are numerous trained approaches that perform phase unwrapping directly, and with noisy inputs [R1, R2]. This paper only compares with DnCNN which is a denoiser for natural images.

 

More comments: 

1. Am I right in understanding that the imaging system is designed to measure changes due to object deformation? If so, is I1 captured while the object is not deformed, while I2 is captued after deformation? How is the deformation controlled?

2. Equations 12, 13 are called Huber loss function, and should be referenced appropriately

3. I am somewhat confused by dataset generation shown in figure 3. Is this paper solving a very specific shearing problem? The various shearing phases look quite similar to me. Why should the shearing be only along x axis?

4. 100 training images is too low. Even with 32x32 patching, the amount of data is not nearly enough to train even a small-sized neural network.

5. Figure 7 does not look like an image that is similar to training set. What is the shear direction here, and how was it generated? Why is the training data not augmented with similar examples? My concern is that the training data is not diverse enough to generalize to real-world settings.

6. The unwrapping results for other approaches are missing.

 

[R1] Wang, Kaiqiang, Ying Li, Qian Kemao, Jianglei Di, and Jianlin Zhao. "One-step robust deep learning phase unwrapping." Optics express 27, no. 10 (2019): 15100-15115.

[R2] Spoorthi, G. E., Rama Krishna Sai Subrahmanyam Gorthi, and Subrahmanyam Gorthi. "PhaseNet 2.0: Phase unwrapping of noisy data based on deep learning approach." IEEE transactions on image processing 29 (2020): 4862-4872.

Comments on the Quality of English Language

Language appears to be satisfactory

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The article "Denoising of wrapped phase in digital speckle shearography based on the convolutional neural network" deals with a current issue that may be helpful in the general signal processing domain. From a practical point of view, it is clear that the phase information of a harmonic signal is essential information that needs to be addressed or given more attention.

This paper tackles the crucial issue of reducing background noise in the extraction of phase information in a given signal/speckle shearography. The authors propose a solution that harnesses the power of CNN, which they term SDCNN in this context. This algorithm holds significant promise for real-world applications, as it can enhance the accuracy and reliability of phase information extraction. 

After studying it, I conclude that the shard has a logical conceptualization, contains all the essential information for the interested reader, and is written in the form of "clear information". The essence of the paper is the implementation of the proposed algorithm for background noise reduction in a given application. The authors build on the theoretical model of DSSPI and present the proposed CNN Denoiser. The following section presents the results, illustrative examples, and their evaluation and application. The authors use an innovative approach to solve the problem, and their results are both valuable, and these results support their conclusions.

After a comprehensive review, I am confident in stating that this article is original and meets all the necessary criteria for publication in our esteemed journal. Based on these findings, I wholeheartedly recommend this article for publication in its current form. 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have satisfactorily answered most of my concerns. The improved manuscript is more informative, and does a fair job comparing against other approaches.