Next Article in Journal
Thermodynamic Jump from Prebiotic Microsystems to Primary Living Cells
Next Article in Special Issue
Images of Roman Imperial Denarii: A Curated Data Set for the Evaluation of Computer Vision Algorithms Applied to Ancient Numismatics, and an Overview of Challenges in the Field
Previous Article in Journal
Organic Molecules: Is It Possible To Distinguish Aromatics From Aliphatics Collected By Space Missions in High-Speed Impacts
Previous Article in Special Issue
Learning to Describe: A New Approach to Computer Vision Based Ancient Coin Analysis
Open AccessArticlePost Publication Peer ReviewVersion 1, Original

Visual Reconstruction of Ancient Coins Using Cycle-Consistent Generative Adversarial Networks (Version 1, Original)

School of Computer Science, University of St Andrews, Scotland KY16 9AJ, UK
Authors to whom correspondence should be addressed.
Received: 8 March 2020 / Accepted: 9 March 2020 / Published: 12 March 2020
(This article belongs to the Special Issue Machine Learning and Vision for Cultural Heritage)
Peer review status: 2nd round review Read review reports

Reviewer 1 Harley R. Myler Lamar University, Department of Electrical Engineering, Beaumont, United States Reviewer 2 Martin Kampel TU Wien Faculty of Informatics Institute of Visual Computing & Human-Centered Technology Computer Vision Lab Reviewer 3 Dimitris Kavroudakis University of the Aegean, Department of Geography, GREECE
Version 1
Authors' response
Not approved
Authors' response
Approved with revisions
Authors' response
Version 2
Reviewer invited
Version 2, Revised
Published: 7 July 2020
DOI: 10.3390/sci2030052
Download Full-text PDF

Version 1, Original
Published: 12 March 2020
DOI: 10.3390/sci2010013
Download Full-text PDF
In this paper, our goal is to perform a virtual restoration of an ancient coin from its image. The present work is the first one to propose this problem, and it is motivated by two key promising applications. The first of these emerges from the recently recognised dependence of automatic image based coin type matching on the condition of the imaged coins; the algorithm introduced herein could be used as a pre-processing step, aimed at overcoming the aforementioned weakness. The second application concerns the utility both to professional and hobby numismatists of being able to visualise and study an ancient coin in a state closer to its original (minted) appearance. To address the conceptual problem at hand, we introduce a framework which comprises a deep learning based method using Generative Adversarial Networks, capable of learning the range of appearance variation of different semantic elements artistically depicted on coins, and a complementary algorithm used to collect, correctly label, and prepare for processing a large numbers of images (here 100,000) of ancient coins needed to facilitate the training of the aforementioned learning method. Empirical evaluation performed on a withheld subset of the data demonstrates extremely promising performance of the proposed methodology and shows that our algorithm correctly learns the spectra of appearance variation across different semantic elements, and despite the enormous variability present reconstructs the missing (damaged) detail while matching the surrounding semantic content and artistic style. View Full-Text
Keywords: deep learning; computer vision; Cycle-GAN; image reconstruction deep learning; computer vision; Cycle-GAN; image reconstruction
Show Figures

Figure 1

MDPI and ACS Style

Zachariou, M.; Dimitriou, N.; Arandjelović, O. Visual Reconstruction of Ancient Coins Using Cycle-Consistent Generative Adversarial Networks. Sci 2020, 2, 13.

Show more citation formats Show less citations formats
Note that from the first issue of 2016, MDPI journals use article numbers instead of page numbers. See further details here.

Article Access Map by Country/Region


Reviewer 1

Sent on 19 Apr 2020 by Harley R. Myler | Approved
Lamar University, Department of Electrical Engineering, Beaumont, United States

This paper was both well-written and well thought out. Certainly an excellent example of the application of GANs with a successful result. I frankly have no negative or constructive comments for you as the paper could readily be used as a tutorial for GANs application!

Response to Reviewer 1

Sent on 11 Jul 2020 by Marios Zachariou, Neofytos Dimitriou, Ognjen Arandjelović

We are most grateful to the reviewer for the kind words - we truly appreciate them! Thank you for your time and effort.

Reviewer 2

Sent on 06 May 2020 by Martin Kampel | Not approved
TU Wien Faculty of Informatics Institute of Visual Computing & Human-Centered Technology Computer Vision Lab

The paper was also submitted to a major European Computer Vision Conference in 2020, which is a violation against the MDPI standards (and might be seen as unethical behaviour):

Manuscripts submitted to MDPI journals should meet the highest standards of publication ethics:

  • Manuscripts should only report results that have not been submitted or published before, even in part.
  • Manuscripts must be original and should not reuse text from another source without appropriate citation.

Response to Reviewer 2

Sent on 11 Jul 2020 by Marios Zachariou, Neofytos Dimitriou, Ognjen Arandjelović

We can assure the reviewer that there was no violation of policy at any stage, though we can see why the reviewer could have thought otherwise (as a side note, to avoid confusion it would have been more collegial to approach the authors directly to ascertain facts before making accusations). In particular we believe that the reviewer may not have understood that this is a post-acceptance review, and not a pre-acceptance one - the process is confusing to us too, no doubt.

Reviewer 3

Sent on 16 Jun 2020 by Dimitris Kavroudakis | Approved with revisions
University of the Aegean, Department of Geography, GREECE

This submission is a very interesting paper with a very applied topic and falls in the scope of the journal. Although, it has some potential for improvement.


  • The general description of the problem (Introduction) and the description of its importance for the science and the society could be further improved.
  • The degree of innovativeness of the methodological approach is not convincingly demonstrated. Some more details about its innovative features could further improve the quality of this paper.
  • Why is this paper likely to be cited in the future?


  • A bit more text regarding the originality of this work and why it contains new results that significantly advance the research field.


  • I believe that adding a bit more text on why the results of the method are satisfactory (evaluation approach) will increase the quality of this work
  • Could the results be more satisfactory if you have changed something in the methodology?
  • Are the results (or the method) sensitive to this specific study area?


  • In the Discussion section I would have wished to see more information on the actual meaning of the findings and how the results add to the broader topic as well as the specific scientific field


  • The "Conclusions" section, could be further improved by describing the importance of this work, the highlight of potential further development of this methodology.

Response to Reviewer 3

Sent on 11 Jul 2020 by Marios Zachariou, Neofytos Dimitriou, Ognjen Arandjelović

We are genuinely very thankful to the reviewer for the positive comments.  Thank you both for these and the constructive feedback which we did our best to address to your satisfaction.  We trust you will find our revision even better, following its improvements effected by revisions in accordance with your feedback.

We do that this is explained and emphasised in the manuscript. In particular, as stated therein (already in the abstract and then reiterated in more detail in the main text) this is the first work to recognize the very possibility of performing virtual reconstruction of ancient coins and the first work to recognize the practical potential in attempting to do this (visualization, pre-processing input to classification methods, etc.), and then naturally the first one to propose a solution to it and demonstrate its promise.

In the text of our manuscript we discuss systematically various design choices, stressing when and (crucially) why these differ from those in the existing literature employing technically related methodologies. We are sure that there always is opportunity for improvement and we certainly hope that our contribution will attract more research in the area, which will no doubt improve on our results.  That being said, if we actually already knew how to do so at this stage, we would have done and reported it. Following the reviewer's other comments, the revision now has more suggestions for future work and possible future improvements, which is perhaps what the reviewer meant here too.

We are not exactly sure what the reviewer is referring to as "this specific study area".  Is the reviewer referring to "ancient numismatics" or "Roman Imperial coins"?  In both cases, and indeed universally so when it comes to the learning of the type adopted in our paper, the answer is yes, as the ability to learn a specific characteristic of data depends on the breath of variation across data, its nature, the practicability of obtaining sufficient training data, etc.  The purposeful domain specific design choices have been highlighted at several stages (e.g. in Section 5.4) with explanations giving the reader further insight as to the phenomena which explain the differential design.  But we do agree that this is worth adding a further note on in the penultimate section and have therefore amended the manuscript accordingly.

We genuinely apologize but we again do not quite understand what the reviewer means by "the actual meaning of the findings".  As we noted when we motivated the work and hence commented upon in the discussion, we aim to create synthetic reconstructions of coins.  The "meaning" is the recovery of missing detail, again as discussed in the paper, and the importance of this is stated already in the manuscript: one can see what the coin looked like before damage took place (visualization) which is important for educational purposes and general appreciation of coins (e.g. by collectors), or indeed use this higher quality synthetic imagery as input to automatic classification and analysis algorithms (again, as stated in the manuscript, including comments on the importance of coin condition in this context with suitable references evidencing this). Nevertheless, to emphasise this further the section has been amdended.

We are genuinely grateful for this suggestion as it made us sit back and reflect further on our work, and come up with actual new ideas! The revision has been amended accordingly - thank you again!

Back to TopTop