Pear Defect Detection Method Based on ResNet and DCGAN

Round 1

Reviewer 1 Report

Please see attachment

Comments for author File: Comments.pdf

Author Response

• The abstract part of the article should put forward some questions, briefly introduce the new method and improvements where it has made. It is suggested that appropriate modifications should be
• The improvement of the method is described in the abstract
• In paragraph (line42-line 47) more references are
• Four related pieces of literature are added in paragraph(line42-line47),
• In paragraph (line48), please explain more your choice for the ResNet (with different variety) compared with other models including the CNN
• Added reasons for choosing ResNet over other models
• You are creating your own pictures for the normal pears and defective pears, so why not taking pictures from different angles and direction as a better way for data augmentation instead the artificial method
• The reasons for adopting the algorithm to achieve image enhancement are explained in paragraph (line107-line115)
• Could you please explain more the parameters in equations 1 and 2 in relation with the real geomorphology of the
• The principle and application of equations 1 and 2 are explained in detail
• The results discussion seems short compared to the insight gained from the application, therefore the work would benefit from a more extensive critical analysis of the results, typical of an insightful research paper as opposed to a technical
• A more detailed description is added in the discussion of experimental results

Author Response File: Author Response.pdf

Reviewer 2 Report

Summary of the paper: in this paper, the authors successfully applied ML methods in pear defect detection, observing better accuracy and efficiency compared to traditional algorithms. The authors first augmented data using both traditional algorithms and DCGAN; by comparing with other models and doing ablation studies, the authors found that ResNet50 with DCGAN is superior to other architectures.

 

The paper has provided enough background and a lot of details for the studies, which makes it qualified for publication in Information; however, there are some critical issues in the paper. Thus, I would recommend publication only after the authors have addressed the following issues:

1. English needs to be improved, which includes but is not limited to: a) grammar errors b) hard-to-read or not understandable sentences c) wrong words or typos. Those problems are not rare in the text.
2. A lot of unclear/undefined terminologies in the paper. For example, f1 to f4 in the caption of Figure 11, DBN in line 36, MCC in line 205, etc.
3. In 4.2, the authors claimed that DCGAN helps models to gain performance improvements compared to models without DCGAN; however, with DCGAN there are more data than without DCGAN. Thus the better performance may come from more data (also means more training). The authors need to give more convincing studies and reasonable analyses to support this statement.
4. In line 112, reinforcement learning is mentioned however it is never actually applied (in fact, reinforcement learning is irrelevant to the work in this paper).
5. Error in Eq (3), in the denominator it should be \sigma instead of \sqrt(\sigma).

Author Response

1. English needs to be improved, which includes but is not limited to: a) grammar errors b) hard-to-read or not understandable sentences c) wrong words or typos. Those problems are not rare in the text.

The wording and grammar are adjusted and typos are corrected.

1. A lot of unclear/undefined terminologies in the paper. For example, f1 to f4 in the caption of Figure 11, DBN in line 36, MCC in line 205, etc.

The terminologies in the paper are explained in detail.

1. In 4.2, the authors claimed that DCGAN helps models to gain performance improvements compared to models without DCGAN; however, with DCGAN there are more data than without DCGAN. Thus the better performance may come from more data (also means more training). The authors need to give more convincing studies and reasonable analyses to support this statement.

The original inaccurate expression is modified. In fact, we want to express that the accuracy of the final result can be improved by using the DCGAN.

1. In line 112, reinforcement learning is mentioned however it is never actually applied (in fact, reinforcement learning is irrelevant to the work in this paper).

The inaccurate expression of the reinforcement learning in line 112 is deleted.

1. Error in Eq (3), in the denominator it should be \sigma instead of \sqrt(\sigma).

The denominator in Eq(3) is modified to \sigma.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

It looks fine to me and the author has answered all questions

Author Response

The textual errors were corrected in detail. And the grammar and wording of the whole paper were also examined.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors have corrected some errors in the paper and also made it more clear now; yet there are still some issues in the paper. Specifically,

• In 4.2, the authors “want to express that the accuracy of the final result can be improved by using the DCGAN”. My main concern is that those accuracy improvements could be due to that the DCGAN generated data is only presented in “With DCGAN” model but not in the “Without DCGAN” model in table 6. So the authors need to clarify that if both models are using the same amount of data. If yes, it would be necessary to make it clear in the text to avoid further confusion; otherwise, authors need to feed the same amount of data into both models to make the statement hold.
• Though the authors have improved the paper, some textual errors still exist in the paper. For example, “performance of the it” in line 260, “he best performance” in line 252, “covering into grayscale images” in line 98, etc.

 

If these two issues are solved, I would recommend it for publication.

Author Response

The textual errors were corrected in detail. And the grammar and wording of the whole paper were also examined.

 

The Cutmix and Mosaic data extending methods were adopted to extend the original data set (the control group without data raising). The other data extending methods also improve the accuracy of models. And the number of processed data set is identical among the three methods. Ultimately, the experimental results expounded DCGAN still has the best performance. Therefore, it is guaranteed that the accuracy improvements by DCGAN do not depend on more data.

Author Response File: Author Response.docx