DRRU-Net: DCT-Coefficient-Learning RRU-Net for Detecting an Image-Splicing Forgery

Round 1

Reviewer 1 Report

This paper presents DDRU-Net that combines RRU-Net and JALM to detect visual artefacts and compression artefacts respectively.

Strength: The highlight of this paper is a new network that is based on existing networks.

Weakness: The objective and authors’ contribution is not clear. It contains mostly implementation details only. Scientific concept of work is completely missing in the paper. Moreover, it is not well structured. For instance, introduction and related work sections contain overlapping information. The latter section contains a lot of details of specific networks rather than a comparison in terms of, e.g., accuracy, speed. It would be better to include a comparison of existing works in form of a Table.

Comments:

·   -      Formatting needs improvements, for example, section numbers are mixed up (there are multiple Section 3.1 and Section 4.2)

·    -     Language requires improvement

·  -       Abstract needs major improvements

·  -       Introduction section needs improvements. For instance, some information from Section 3.1 (page 6) can be added

·  -       Section 3.1 (page 4): architecture requires more details, e.g., JALM

·   -      Table 2: which image belongs to CASIAv2?

 

Author Response

Thank you for your thoughtful review.

Comments:

1. Formatting needs improvements, for example, section numbers are mixed up (there are multiple Section 3.1 and Section 4.2)
--> All relevant parts have been corrected.

2.  Language requires improvement
--> Fully proofread by native speakers.

3. Abstract needs major improvements
--> The abstract was rewritten to make it easier to understand the gist of this paper.

4.  Introduction section needs improvements. For instance, some information from Section 3.1 (page 6) can be added
--> The introduction was modified by adding the contents of 3.1.

5. Section 3.1 (page 4): architecture requires more details, e.g., JALM
--> Added pictures and content for JALM architecture and procedure.

6. Table 2: which image belongs to CASIAv2?
--> Added DEFACTO image and CASIAv2 image labels to Table 2.

 

Author Response File: Author Response.docx

Reviewer 2 Report

This manuscript proposed the DRRU-Net for detecting image splicing forgery. Experimental results prove the effectiveness of the proposed approach. However, some revisions are needed to improve the manuscript.

1.      The proposed approach should have some mathematical explanation.

2.      It is recommended to add some more references from the past three years and compare them.

 

3.      The size of the figures in the manuscript needs to be adjusted appropriately.

Author Response

Thank you for your thoughtful review.

Comments:
This manuscript proposed the DRRU-Net for detecting image splicing forgery. Experimental results prove the effectiveness of the proposed approach. However, some revisions are needed to improve the manuscript.

1. The proposed approach should have some mathematical explanation.
--> The structure of DRRU-Net for detecting image splicing forgery was explained through Equations 1 to 7, and the evaluation method for verifying the performance of DRRU-Net was described through the remaining equations.

2. It is recommended to add some more references from the past three years and compare them.
--> Added literature on recent relevant studies. [8-10], [13-15], [18-20]
 

3. The size of the figures in the manuscript needs to be adjusted appropriately.
--> The picture for image forgery detection is intentionally enlarged so that comparison can be made with the naked eye. This part will be modified upon request during the editing process.

Author Response File: Author Response.docx

Reviewer 3 Report

This manuscript applsci-2200734 proposed a lightweight deep learning network for detecting an image splicing forgery. The research on image forgery detection using CNN, a deep learning network, and research on detecting and localizing forgery in pixel units are in progress. Among them, CAT-Net, which learns the discrete cosine transform (DCT) coefficients of images together with images, was released in 2022. The DCT coefficients presented by CAT-Net are combined with the learning module (JPEG Artifact Learning Module) and the backbone model as pre-learning, and the weights are fixed. The dataset used for pre-training is not included in the public dataset, and the backbone model has a relatively large number of network parameters, which causes overfitting in a small dataset, hindering generalization performance. In this paper, this learning module is designed to learn the characterization depending on the DCT domain in real-time during network training without pre-training. The DRRU-Net proposed in this paper is a network that combines RRU-Net which detects forgery by learning only images and JPEG artifact learning module. It was a pleasure reviewing this work and I can recommend it for publication in Applied Science after a major revision. I respectfully refer the authors to my comments below.

 

1.       The English needs to be revised throughout. The authors should pay attention to the spelling and grammar throughout this work. I would only respectfully recommend that the authors perform this revision or seek the help of someone who can aid the authors.

2.       (Reference) Please check the style of all the references. Replace the references from the arXiv with the published papers, such as [2] and [21]. Because the manuscripts from arXiv are not the published works.

3.       (Page 3) In the Introduction part, “main contributions” is best to list clearly by breaking it down into several points.

4.       Overall, it is necessary to unify the 'JPEG learning module' and the 'JPEG arifact learning module' into one expression. (e.g. Fig. 2 uses 'JPEG learning module')

5.       It would be better to express 'Out Conv' in Fig 2. as a full name.

6.       (L.177) typo error - ','

7.       (L.483) typo error

8.       It would be nice to add references to the DEFACTO and CASIAv2 datasets used in the experiments in Chapter 4.

My overall impression of this manuscript is that it is in general well-organized. The work seems interesting and the technical contributions are solid. I would like to check the revised manuscript again.

Author Response

Thank you for your thoughtful review.
 
Comments:

1. The English needs to be revised throughout. The authors should pay attention to the spelling and grammar throughout this work. I would only respectfully recommend that the authors perform this revision or seek the help of someone who can aid the authors.
--> Fully proofread by native speakers.

2. (Reference) Please check the style of all the references. Replace the references from the arXiv with the published papers, such as [2] and [21]. Because the manuscripts from arXiv are not the published works.
--> Checked and corrected the applicable references and also fixed all malformed parts of the references.

3. (Page 3) In the Introduction part, “main contributions” is best to list clearly by breaking it down into several points.
--> The main contribution is specified in lines 94 to 111.

4. Overall, it is necessary to unify the 'JPEG learning module' and the 'JPEG arifact learning module' into one expression. (e.g. Fig. 2 uses 'JPEG learning module')
--> Corrected that part.

5. It would be better to express 'Out Conv' in Fig 2. as a full name.
--> Corrected that part.

6. (L.177) typo error - ','
--> Corrected that part.

7. (L.483) typo error
--> Corrected that part.

8. It would be nice to add references to the DEFACTO and CASIAv2 datasets used in the experiments in Chapter 4.
--> Added references [31,32] for each dataset.

My overall impression of this manuscript is that it is in general well-organized. The work seems interesting and the technical contributions are solid. I would like to check the revised manuscript again.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Related work section contains a lot of details of specific networks rather than a comparison with the new approach, for example in terms of accuracy. In the previous review, it was suggested that the authors should improve this section, for example by adding an detailed comparison with existing works. However, this is not sufficiently improved so far. On page 3, now some details are included but they are still mostly implementation details. Both RRU-Net and CAT-net are existing networks.

Author Response

Thank you for your thoughtful review again.

Comments:
Related work section contains a lot of details of specific networks rather than a comparison with the new approach, for example in terms of accuracy. In the previous review, it was suggested that the authors should improve this section, for example by adding an detailed comparison with existing works. However, this is not sufficiently improved so far. On page 3, now some details are included but they are still mostly implementation details. Both RRU-Net and CAT-net are existing networks.

--> In Chapter 2 Background, the differences between RRU-Net and CAT-Net, which are existing studies, and DRRU-Net proposed in this paper are added. (L.161-167, L.184-190)

 

Round 3

Reviewer 1 Report

Concerns have been addressed