Two-Stage Robust Lossless DWI Watermarking Based on Transformer Networks in the Wavelet Domain

Round 1

Reviewer 1 Report

Summary: In present paper, the authors studied Two-Stage Robust Lossless DWI Watermarking Based on Transformer Networks in Wavelet Domain.

Evaluation: Authors have prepared a well-discussed work which presents a well-considered up-to-date aspect which has interest for all. The novelty of the idea of the authors is obvious. Presented algorithm and analysis are correct. I recommend this article to be published in Applied Sciences, but it needs the following minor corrections:

1.   Authors should check and correct the typological mistakes sincerely throughout the paper.

2.     In the introduction section, add the motivation and the importance of the considering the present study over the existing studies.

3.     Polish merits to your proposed method and what are limitations of the method.

4.     Add some gaps between table and its caption. Elaborate all Figures in the tex.

5.     On page 13, Check BER values in Table 3.

6.     Irrelevant references should be deleted from the list of references.

7.     Authors should add the conclusion section with focus on both impact and insights of the manuscript. Clearly state your unique research contributions in the conclusion section and point out some potential directions for future work.

Moderate editing of English language.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

In the manuscript, the authors proposed a two-stage transformer network for robust lossless watermarking, in order to solve inevitable irreversible distortion to diffusion-weighted imaging (DWI) images caused by existing robust watermarking algorithms.

In the first two sections, the authors briefly introduced the mechanism of DWI technique, the development of digital watermarking, highlighted its necessity, and pointed out the limit of current robust watermarking algorithms.

The network is mainly composed of three parts, an encoder to reconstruct the watermarked image, an attack layer network to attack the watermarked image through conventional image and geometric attack methods, and a decoder to extract the watermark information from the distorted image from the attack layer network. There is also a separate module for reversible information embedding.

In terms of the training, the first stage is to train the robust watermarking network with the frequency information enhancement module to improve the reconstruction quality. The second stage is to produce the watermarked image via reversible embedding, based on the pre-trained robust watermarking network. Finally, the watermark extraction network is trained on the second embedding result to avoid weakening the robustness of the first stage caused by the reversible embedding. The authors provided sufficient details for the audience to understand the architecture of each part of the network and the training process.

In the experiment, Bit Error Rate (BER) and Peak Signal-to-Noise Ratio (PSNR) were used to evaluate the robustness of the algorithm. The results shown are overall promising. However, I think there are a few points the authors can consider to further improve the quality of the manuscript:

1.     For each component in the network, add a typical output figure to help the audience better understand its functionality. Same as Figure 9, in the experiment result discussion the authors should also provide typical image output from different components.

2.     Without providing the source code, considering the complexity of the whole network, at least the authors should provide information of general model size of each component to obtain the performance shown in the experiment.

3.     The authors should provide more details of the training history.

 

4.     I don’t understand why MSE loss between I_0 and I_e is used in the training of the first stage. With watermark image information embedded, there is no way for the encoded image to be the same as the original image.

Try to simplify sentences for a better reading experience.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The suggested Lossless DWI Watermarking work must be substantially reworked on its write-up and achieve results.

1. At present, the research gap and objective of the work from the abstract is not clear.

2. The PSNR of the watermarked image reaches 60.18dB. At what embedding capacity? How much is the hiding capacity?

3. Nowhere the authors mentioned the research objectives: Is it to preserve file size, or to maintain visual quality, or anything else? It has to be mentioned in a separate section with an in-depth discussion of the existing research issues.

4. The proposed work has to be explained through a working example step by step clearly and concisely so that any reader who looks at the paper can understand it.

5. A numerical example needs to explain Two-level Haar wavelet decomposition.

6.  The authors are suggested to refer to the recent works with in last three years to perform the watermark analysis to various intentional and un-intentional attacks, such as the work presented in below papers needs to discussed with its merits and issues: https://doi.org/10.3390/electronics12051222, https://doi.org/10.1007/s12652-021-03365-9  

7. A numerical illustration of the embedding and extraction process should be presented.

8. At least 50-100 images must be considered for validating the obtained result.

9. Mention the limitation and implications of the work.

 

 

 

Quality of English Language is good

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Authors have revised the manuscript sucessfully

Minor editing of the English language required