Data Enhancement for Plant Disease Classification Using Generated Lesions

Round 1

Reviewer 1 Report

1. The grammatical quality of the manuscript must be greatly enhanced as well as amending colloquial expressions and numerous careless mistakes (typo, wrong citation, wrong legend, etc.). More importantly, the current scientific logical structure provides the reviewer extensive difficulty (lack of: sufficient background information, the problem the authors want to solve, scientific significance). Points above must be improved to make possible a deep review. In my opinion, manuscript reconstruction post English proofreading and a third person check is mandatory.

2. Your idea of generating plant disease images by generating lesion over leaves via GAN, moreover combining with smoothing to make the image naturally is very interesting. The proposed approach itself seems to contain sufficient scientific significance.

3. The name Binarized GAN is misleading, as there are already an architecture named binarized CNN that does not utilize floats. Your network is rather related to GAN usage of "Image Completion" or "Inpainting", or more in detail, compositional GANs. Consider renaming referring to related literatures.

4. If you want to emphasize your Network performance comparing with other benchmark networks, that should be done prior to image smoothing, or else it will not be a fair comparison. If you intend to emphasize the pipeline of a set of GAN and smoothing, you should express in the manuscript a "pipeline" or "approach" rather than the GAN name itself.

 

Author Response

Dear Reviewer,

    Thanks very much for taking your time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses in below and my revisions/corrections in the resubmitted files.

 

YOUR SUGGESTIONS:

The grammatical quality of the manuscript must be greatly enhanced as well as amending colloquial expressions and numerous careless mistakes (typo, wrong citation, wrong legend, etc.). More importantly, the current scientific logical structure provides the reviewer extensive difficulty (lack of: sufficient background information, the problem the authors want to solve, scientific significance). Points above must be improved to make possible a deep review. In my opinion, manuscript reconstruction post English proofreading and a third person check is mandatory. Your idea of generating plant disease images by generating lesion over leaves via GAN, moreover combining with smoothing to make the image naturally is very interesting. The proposed approach itself seems to contain sufficient scientific significance. The name Binarized GAN is misleading, as there are already an architecture named binarized CNN that does not utilize floats. Your network is rather related to GAN usage of "Image Completion" or "Inpainting", or more in detail, compositional GANs. Consider renaming referring to related literatures. If you want to emphasize your Network performance comparing with other benchmark networks, that should be done prior to image smoothing, or else it will not be a fair comparison. If you intend to emphasize the pipeline of a set of GAN and smoothing, you should express in the manuscript a "pipeline" or "approach" rather than the GAN name itself.

MY RESPONSE:

Compared to the first version, we have marked red in modified section.

 

Reviewer #1: The grammatical quality of the manuscript must be greatly enhanced as well as amending colloquial expressions and numerous careless mistakes (typo, wrong citation, wrong legend, etc.).

Response #1: Thank you very much for your suggestion, and we apologize for our previous carelessness. We have changed the grammar as much as possible. At the same time, we also invited MDPI English Editer to give guidance on grammar.

 

Reviewer #2: More importantly, the current scientific logical structure provides the reviewer extensive difficulty (lack of: sufficient background information)

Response #2: Based on your suggestions, we have made many changes to the paper, and the sufficient background information has been added on line [23-31]. It emphasizes how much economic losses can be caused by plant diseases and even lead to disasters such as species extinction. And if we can effectively deal with it at an early stage, we can avoid huge losses. The use of neural networks can enable people without relevant backgrounds to accurately identify the corresponding disease, which can greatly shorten the time to identify the disease, thereby reducing losses(line [30-36]).

 

Reviewer #3: lack of: the problem the authors want to solve, scientific significance

Response #3: We divide the problem we are solving into two types, one is of scientific significance: added in line [48-51](What scientific research can our method be used for), and the other is of advantages of our method(what problems can be solved by using binarized image and GAN in the research of plant lesions.): added in line [111-117]. And we added the GAN to generate a complete Pepper bell Bacterial spot lesion image and its Inception score and Fréchet Inception Distance evaluation in Appendix A [349] to confirm our scientific significance.

 

Reviewer #4: The name Binarized GAN is misleading, as there are already an architecture named binarized CNN that does not utilize floats. Your network is rather related to GAN usage of "Image Completion" or "Inpainting", or more in detail, compositional GANs. Consider renaming referring to related literatures.

Response #4: I am sorry that we have not seriously considered this problem before. We combined your Reviewer #5's suggestion and changed our method name to Binarization generation network combined with image smoothing (called: ES-BGNet)[51-52]. This is the result of our discussion. We think that this can describe our method in more detail, while avoiding the misleading problem of the name.

 

Reviewer #5: If you want to emphasize your Network performance comparing with other benchmark networks, that should be done prior to image smoothing, or else it will not be a fair comparison. If you intend to emphasize the pipeline of a set of GAN and smoothing, you should express in the manuscript a "pipeline" or "approach" rather than the GAN name itself.

Response #5: We are aware of the problem you are talking about. We have made extensive changes to the paper. In addition, we have deleted the comparison with DCGAN and WGAN-GP. The purpose of our paper is to propose a method for generating plant lesion data instead of a new GAN. we added a lot of descriptions of our method (such as line [11-12], line [51], line [328]). At the same time, the scientific significance of Response #3 also emphasizes that our purpose is to propose a method. In order to prove that our method improves the recognition accuracy of classification networks. We invited plant disease experts and used Alexnet to classify the generated lesions and real lesions and generate images to augment the accuracy of Alexnet training to verify our generation effect (line [296-326]).

 

Thank you again for your valuable suggestions, and I will continue to make progress.

Have a nice day.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposes a binary generative models for plant lesion leave generation. The idea is interesting but there are several concerns need to be addressed:

Binary GAN has been proposed and studied previously by several works, and it is not a new idea. Ideally more technical novelty is expected from this work. There should be a discussion on what is the different and advantages of the binary generative model in this work, versus the previous binary GAN models, such as [a] and [b] below. The notation for some of the equations is not aligned, and should be fixed, such as Eq (2). In the experimental results, the model is compared with some of the older versions of GAN, such as vanilla GAN and W-GAN. There has been several improvement of GAN over the past two years, and author need to make a comparison with more recent works too. There should be comparison with previous Binary GAN models, such as [a] and [b] below: Are the accuracy rates reported in Table 2 sample dependent? Given the limited number of samples, they should report sample independent accuracy. What is the confidence interval of these accuracy rates? Are the improvement significant? There are several grammatical errors and typos in this work. The authors should proofread this work, and improve its English written significantly. Can the authors benefit from using attention mechanism to create more realistic lesions? Using attention in CNN is shown to help in several generative tasks, by attending to the more salient part of the data. Many of the relevant works on applying GAN for data augmentation and synthetic data generation, as well as attention based models, are missing from the references.

[a] Binary Generative Adversarial Networks for Image Retrieval

[b] Training Generative Adversarial Networks with Binary Neurons by End-to-end Backpropagation

[d] Deep-Emotion: Facial Expression Recognition Using Attentional Convolutional Network

[d] Self-Attention Generative Adversarial Networks

Author Response

Dear Reviewer,

    Thanks very much for taking your time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses in below and my revisions/corrections in the resubmitted files.

 

YOUR SUGGESTIONS:

This paper proposes a binary generative models for plant lesion leave generation. The idea is interesting but there are several concerns need to be addressed:

Binary GAN has been proposed and studied previously by several works, and it is not a new idea. Ideally more technical novelty is expected from this work. There should be a discussion on what is the different and advantages of the binary generative model in this work, versus the previous binary GAN models, such as [a] and [b] below. The notation for some of the equations is not aligned, and should be fixed, such as Eq (2). In the experimental results, the model is compared with some of the older versions of GAN, such as vanilla GAN and W-GAN. There has been several improvement of GAN over the past two years, and author need to make a comparison with more recent works too. There should be comparison with previous Binary GAN models, such as [a] and [b] below: Are the accuracy rates reported in Table 2 sample dependent? Given the limited number of samples, they should report sample independent accuracy. What is the confidence interval of these accuracy rates? Are the improvement significant? There are several grammatical errors and typos in this work. The authors should proofread this work, and improve its English written significantly. Can the authors benefit from using attention mechanism to create more realistic lesions? Using attention in CNN is shown to help in several generative tasks, by attending to the more salient part of the data. Many of the relevant works on applying GAN for data augmentation and synthetic data generation, as well as attention based models, are missing from the references.

[a] Binary Generative Adversarial Networks for Image Retrieval

[b] Training Generative Adversarial Networks with Binary Neurons by End-to-end Backpropagation

[c] Deep-Emotion: Facial Expression Recognition Using Attentional Convolutional Network

[d] Self-Attention Generative Adversarial Networks

MY RESPONSE:

Compared to the first version, we have marked red in modified section.

 

Reviewer #1: There should be a discussion on what is the different and advantages of the binary generative model in this work, versus the previous binary GAN models, such as [a] and [b] below.

Response #1: Based on your suggestion, we have added the advantages and implementation differences of our method in line [111-117] (Compared to other methods, through attention and GAN, we can solve the problem that the generated lesion images with randomly shape must to be manually marked to synthesize with the leaves, resulting in a waste of time and cost.). And we added the GAN to generate a complete Pepper bell Bacterial spot lesion image and its Inception score and Fréchet Inception Distance evaluation in Appendix A [349] to confirm our approach.

 

Reviewer #2: The notation for some of the equations is not aligned, and should be fixed, such as Eq (2).

Response #2: I am sorry we didn't realize the problem before. In the new paper, We have copied the formula template provided by appli-sci for every formula to solve the problem of misalignment of the formula.

 

Reviewer #3: In the experimental results, the model is compared with some of the older versions of GAN, such as vanilla GAN and W-GAN. There has been several improvement of GAN over the past two years, and author need to make a comparison with more recent works too.

Response #3: We have changed the paper based on your Reviewer #1 suggestion. We think our purpose of the paper is to propose a method for generating plant lesion data instead of a new GAN. So we will The comparison experiments between GAN and DCGAN and WGAN-GP were deleted, and explained our method at line [11-12, 51, 328]. We invited plant disease experts and used Alexnet to classify the generated lesions and real lesions, and generate images to augment the accuracy of Alexnet training to verify our generation effect. As for the choice of GAN comparison, we refer to the GAN comparison carried out in [e] (line [111]), and selected the loss of WGAN-GP as the loss function of our network.

 

Reviewer #4: There should be comparison with previous Binary GAN models, such as [a] and [b] below

[a] Binary Generative Adversarial Networks for Image Retrieval

[b] Training Generative Adversarial Networks with Binary Neurons by End-to-end Backpropagation

Response #4: Thank you very much for your suggestions. We have thought and debated carefully. We have cited the references you mentioned, and in the latest paper we emphasized that we are proposing a method to augment the lack of plant disease spot data, not a new GAN (line [11-12, 51, 328]). Because our method solves the problem of enhancing plant lesion leaf data that requires leaf information and lesion information to be accurately identified and it is difficult to obtain corresponding data due to infectivity and other reasons. There are many such plant diseases, such as citrus canker and pitaya canker (line [45-46]). We also compared the recognition accuracy of Alexnet with or without leaf information in Table A2 of Appendix A (line [369]). About [a] and [b] and other new references we added, due to their methods solve different problems from us. For example, [a] uses GANs to generate images into binary data and restored to the original image. [b] is a training method for improving GAN using a binary method. None of their methods can produce a plant lesion leaf with a specific shape. This led to our method used edge smoothing algorithms, but these binary methods did not use edge smoothing. We think whether this will lead to unfair contrast. We are also continuing to carry out more comparative experiments and improvements to get better results.

 

Reviewer #5: Are the accuracy rates reported in Table 2 sample dependent? Given the limited number of samples, they should report sample independent accuracy.

Response #5: Thank you very much for your suggestions. In the new paper,  we calculated the Pearson correlation coefficients for the samples in Table 2 (line [320]), Table A2 (line [369]), and Table A3 (line [375]). It confirms that our sample accuracy is independent. The samples we use are the PlantVillage public dataset and our own Citrus canker dataset taken with the Nikon D7500. At the same time, we also used the pitaya canker dataset found on the network in the comparison experiment (Table A2, line[369]).

 

Reviewer #6: What is the confidence interval of these accuracy rates? Are the improvement significant?

Response #6: Based on your suggestion, we calculated the P value of the accuracy in 10 iterations, and the confidence interval at P <0.05. It is confirmed that the sample plant lesion data expanded by our method can indeed significantly improve the recognition accuracy of Alexnet when added to Alexnet training (Table 2, line [320]). At the same time, to make the paper more convincing, we calculated the corresponding confidence intervals in each table (Table A1, line[365]; Table A2, line[369]; Table A3, line [375]).

 

Reviewer #7: There are several grammatical errors and typos in this work. The authors should proofread this work, and improve its English written significantly.

Response #7: Thank you very much for your suggestion, and we apologize for our previous carelessness. We have changed the grammar as much as possible. At the same time, we also invited MDPI English Editer to give guidance on grammar.

 

Reviewer #8: Using attention in CNN is shown to help in several generative tasks, by attending to the more salient part of the data.

Response #8: Based on your Reviewer #1 suggestion, we have increased the advantages and differences of our method (line [111-117]). This is a good illustration that using attention in CNN has been shown to help in several generative tasks. At the same time, they are tested for independence and calculated with confidence intervals. It can be seen that our method does benefit from the attention.

 

Reviewer #9: Many of the relevant works on applying GAN for data augmentation and synthetic data generation, as well as attention based models, are missing from the references.

Response #9: In the newly revised paper, we removed some references that were not relevant to our paper. At the same time, on the basis of [a], [b], [c], [d] you suggested, we also found many references related to our method, also used GAN to generate plant leaf data and improved the relevant network performance, or also uses attention to solve the corresponding problems of GANs (line [94-110]). We will continue to read more paper to improve the performance of our method.

 

Thank you again for your valuable suggestions, and I will continue to make progress.

Have a nice day.

 

[a] Binary Generative Adversarial Networks for Image Retrieval

[b] Training Generative Adversarial Networks with Binary Neurons by End-to-end Backpropagation

[c] Deep-Emotion: Facial Expression Recognition Using Attentional Convolutional Network

[d] Self-Attention Generative Adversarial Networks

[e] Lucic M, Kurach K, Michalski M, et al. Are gans created equal? a large-scale study[C]//Advances in neural information processing systems. 2018: 700-709.

Author Response File: Author Response.docx

Reviewer 3 Report

From reading the article, it is not clear why deep learning is required. The author should have done evaluation to show what's the performance that conventional machine learning can obtain, especially considering that one of the referenced article can get really high accuracy with Support Vector Machine. Furthermore, the author doesn't clearly mentioned what is the characteristic of a canker in a plant. The author said that initially the image generated was not good but the author didn't show it and say why they are not good. What is the evaluation metric to say the image generated is good/bad. Have you tested on just noise data?

There are a lot of grammar mistakes making the article hard to read.

Author Response

Dear Reviewer,

    Thanks very much for taking your time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses in below and my revisions/corrections in the resubmitted files.

 

YOUR SUGGESTIONS:

From reading the article, it is not clear why deep learning is required. The author should have done evaluation to show what's the performance that conventional machine learning can obtain, especially considering that one of the referenced article can get really high accuracy with Support Vector Machine. Furthermore, the author doesn't clearly mentioned what is the characteristic of a canker in a plant. The author said that initially the image generated was not good but the author didn't show it and say why they are not good. What is the evaluation metric to say the image generated is good/bad. Have you tested on just noise data?

There are a lot of grammar mistakes making the article hard to read.

MY RESPONSE:

Compared to the first version, we have marked red in modified section.

 

Reviewer #1: From reading the article, it is not clear why deep learning is required.

Response #1: I am sorry for not explaining this enough before. Due to the complex structure of plant lesions, traditional computer vision techniques seek to program the model to identify a series of traits that empirically. This led to the diagnosis of plant lesion was difficult. The neural network learn these complex structural information by itself, so we use the neural network as our evaluation method. We explained in line [32-36, 73-75] of the most recently submitted article.

 

Reviewer #2: The author should have done evaluation to show what's the performance that conventional machine learning can obtain, especially considering that one of the referenced article can get really high accuracy with Support Vector Machine.

Response #2: This is our mistake. We skipped this step before and used the neural network to classify the plant lesion image directly. Based on your comments, in Table A3 of Appdix A (line [375]), we compared the recognition results of several commonly used machine learning methods and neural networks. And by reading related literature, we find that traditional methods require complex image preprocessing and the recognition effect is not as accurate as neural networks.

 

Reviewer #3: Furthermore, the author doesn't clearly mentioned what is the characteristic of a canker in a plant.

Response #3: In the latest paper, we explained that the structure of plant lesions is complex, the colors are diverse, and many lesions are very similar, which makes it difficult to accurately describe the structure of lesions, so we use neural networks (line [33-34]). At the same time, your three suggestions, Reviewer # 1, Reviewer # 2, and Reviewer # 3, made me realize that the previous article was not rigorous enough. In the latest article, we have made extensive modifications to clearly explain the problem we solved (line [48-51]), our advantages (line [111-117]) to ensure the rigor of the paper.

 

Reviewer #4: The author said that initially the image generated was not good but the author didn't show it and say why they are not good.

Response #4: In Figure A1 of Appendix A (line [359]), we added the GAN to generate the complete Citrus canker and Pepper bell Bacterial spot image. In Table A1 of Appendix A (line [365]), we added the corresponding Inception score and Fréchet Inception Distance.  As can be seen, the effect of generating intact diseased leaves is very poor. At the same time, we carefully checked the part of the article that needs to be verified, and added a comparative experiment on the accuracy of the lesion recognition accuracy with or without leaf information when the lesions are very similar (Appendix A, Table A2 line [369]). The comparative dataset was citrus canker and pitaya canker. Citrus canker and pitaya canker are shown in Figure 1 (line [60]).

 

Reviewer #5: What is the evaluation metric to say the image generated is good/bad.

Response #5: In the latest article, we use human expert and Alexnet discrimination(line [296-301]), and Alexnet recognition accuracy improvement (line [302-326]) results for evaluation. Because we use an edge smoothing algorithm, but other GANs cannot add edge smoothing algorithms, this is unfair. And in the latest paper, we mainly explained that our paper is to propose a method for generating complete plant lesion data (line [11-12, 51, 328]), rather than a new GAN method. Inception score and Fréchet Inception Distance are only used for complete generated image quality assessment in Appendix A. At the same time, in Table A1 of Appendix A (line [365]), we have added the comparison of the Inception Score and Fréchet Inception Distance when using WGAN-GP and DCGAN to generate a complete plant lesion leaf, because this is a fair comparison.

 

Reviewer #6: Have you tested on just noise data?

Response #6: The generated image we tested is an image after IS stabilization, and randomly select specified number (such as 500 generated images) that we want to expand the dataset; In each dataset selection, we ensure sufficient fairness to reduce errors caused by human factors.

 

Reviewer #7: There are a lot of grammar mistakes making the article hard to read.

Response #7: Thank you very much for your suggestion, and we apologize for our previous carelessness. We have changed the grammar as much as possible. At the same time, we also invited MDPI English Editer to give guidance on grammar.

 

Thank you again for your valuable suggestions, and I will continue to make progress.

Have a nice day.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I appreciate the authors’ extensive effort in the revision of the manuscript in response to the reviewer’s comment. Although I personally feel there is still room for improvement of the manuscript in providing a further experimental result for reinforcing the authors’ proposed method, the minimum requirement seems to be stated. I feel readers can interpret and understand your scientific progress with the current information provided. 

I do not need a further response nor revisions, but please amend or discuss the following point with the editor.

1. No description is provided to the suppl. dataset so this is my presumption however, your supplemental data includes raw data that seems to be derived from other authors dataset (e.g. Plantvillage). The authors should take extensive care NOT TO violate or infringe the original copyright holder. If you have permissions from them to redistribute the original data, state in the manuscript. Moreover, even if you have permission from them, I feel the supplemental data provides no information to the reader so should be removed in the current form.

2. Discuss with the editor whether program codes to reproduce the results should be provided suppose in github repository. I personally believe it is mandatory to enable a third person to reproduce and validate your results, however, this depends to the journal policy.

Author Response

Dear Reviewer,

    I am sorry, you mentioned that you do not need a further response nor revisions, but we must reply you before submit the revised manuscript on the website. Otherwise we cannot submit a revised paper.

    Thanks very much for taking your time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses in below and my revisions/corrections in the resubmitted files.

 

YOUR SUGGESTIONS:

 

I appreciate the authors’ extensive effort in the revision of the manuscript in response to the reviewer’s comment. Although I personally feel there is still room for improvement of the manuscript in providing a further experimental result for reinforcing the authors’ proposed method, the minimum requirement seems to be stated. I feel readers can interpret and understand your scientific progress with the current information provided.

 

I do not need a further response nor revisions, but please amend or discuss the following point with the editor.

 

No description is provided to the suppl. dataset so this is my presumption however, your supplemental data includes raw data that seems to be derived from other authors dataset (e.g. Plantvillage). The authors should take extensive care NOT TO violate or infringe the original copyright holder. If you have permissions from them to redistribute the original data, state in the manuscript. Moreover, even if you have permission from them, I feel the supplemental data provides no information to the reader so should be removed in the current form.

 

Discuss with the editor whether program codes to reproduce the results should be provided suppose in github repository. I personally believe it is mandatory to enable a third person to reproduce and validate your results, however, this depends to the journal policy.

 

MY RESPONSE:

 

Compared to the first version, we have marked yellow in modified section.

 

Reviewer #1: No description is provided to the suppl. dataset so this is my presumption however, your supplemental data includes raw data that seems to be derived from other authors dataset (e.g. Plantvillage). The authors should take extensive care NOT TO violate or infringe the original copyright holder. If you have permissions from them to redistribute the original data, state in the manuscript. Moreover, even if you have permission from them, I feel the supplemental data provides no information to the reader so should be removed in the current form.

 

Response #1: Thank you very much for your suggestions. The Citrus canker and pitaya dataset in our dataset were collected for our own shooting. The PlantVillage dataset provided by the organization kaggle. With reference to kaggle's Privacy Policy, we are indeed suspected of infringement. With reference to your suggestions, we have deleted the PlantVillage data comparison section.

Sincerely thank you for your valuable reminder, so that we have avoided one possible legal dispute in the future.

 

Reviewer #2: Discuss with the editor whether program codes to reproduce the results should be provided suppose in github repository. I personally believe it is mandatory to enable a third person to reproduce and validate your results, however, this depends to the journal policy.

 

Response #2: Based on your suggestions. In the latest paper, we put the code corresponding to the paper in https://github.com/Ronzhen/ES-BGNet (line [252]).The time required for the algorithm to run 100,000 times on the GTX 1070Ti is also mentioned in the paper (line [252,307]). To provide readers with more convenience. We also hope that our paper can be applied in practice to solve a certain problem. At the same time, we are also making a public platform, hoping to use the data we collected for more scientific research or practical work.

 

Thank you again for your valuable suggestions, because of your suggestion, I also gained a lot in this paper work. We will work harder in future research work to improve our ability in scientific research, English, writing and so on.

Have a nice day.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper has significantly been changed compared to the previous version. In the previous version the main focus was on binary GAN, and in this version it seems that the focus has shifted to the application of synthesized sample generated by GANs for training a model in the presence of limited labeled samples. The authors have addressed some of my concerns, and I believe they should address the following concerns too, to be ready for publication:

 

The choice of model parameters/hyperparameters should be better explained. For example in Eq (6) why is lambda set to 0.2. Is it based on experimental results on generative model? Or accuracy? Or is it based on some previous study? The same for bi-linear interpolation, what bi-linear? Adding some discussion to justify these choices can help reader to better understand these choices. There are still some grammatical error in this paper, and these need to be fixed. In Table A.1, it would be great to add the IS and FIS of some of the more recent GAN models too, beside WGAN and DCGAN. A more detailed discussion on the training/test time of this approach would be helpful for readers who want to follow a similar idea. Since the main idea of this paper is to use GAN for data synthesis, the authors should add more recent works which use this idea for data augmentation. This idea has been heavily used for image classification, person re-identification, and various biometrics such as fingerprint, iris, and face. The following papers have some of the relevant works:

[a] "Unlabeled samples generated by gan improve the person re-identification baseline in vitro." Proceedings of the IEEE International Conference on Computer Vision. 2017.

[b] “Biometric Recognition Using Deep Learning: A Survey”. arXiv preprint arXiv:1912.00271. 2019.

Author Response

Dear Reviewer,

    Thanks very much for taking your time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses in below and my revisions/corrections in the resubmitted files.

 

YOUR SUGGESTIONS:

This paper has significantly been changed compared to the previous version. In the previous version the main focus was on binary GAN, and in this version it seems that the focus has shifted to the application of synthesized sample generated by GANs for training a model in the presence of limited labeled samples. The authors have addressed some of my concerns, and I believe they should address the following concerns too, to be ready for publication:

 

The choice of model parameters/hyperparameters should be better explained. For example in Eq (6) why is lambda set to 0.2. Is it based on experimental results on generative model? Or accuracy? Or is it based on some previous study? The same for bi-linear interpolation, what bi-linear? Adding some discussion to justify these choices can help reader to better understand these choices. There are still some grammatical error in this paper, and these need to be fixed. In Table A.1, it would be great to add the IS and FIS of some of the more recent GAN models too, beside WGAN and DCGAN. A more detailed discussion on the training/test time of this approach would be helpful for readers who want to follow a similar idea. Since the main idea of this paper is to use GAN for data synthesis, the authors should add more recent works which use this idea for data augmentation. This idea has been heavily used for image classification, person re-identification, and various biometrics such as fingerprint, iris, and face. The following papers have some of the relevant works:

[a] "Unlabeled samples generated by gan improve the person re-identification baseline in vitro." Proceedings of the IEEE International Conference on Computer Vision. 2017.

 

[b] "Biometric Recognition Using Deep Learning: A Survey". arXiv preprint arXiv:1912.00271. 2019.

 

MY RESPONSE:

Compared to the first version, we have marked yellow in modified section. At the same time, because the pepper bell bacterial spot dataset is very similar to the experiment of the Citrus canker dataset, and because we have infringements using the PlantVillage dataset. We removed experiments related to pepper bell bacterial spot dataset in the revised paper.

 

Reviewer #1: The choice of model parameters/hyperparameters should be better explained. For example in Eq (6) why is lambda set to 0.2. Is it based on experimental results on generative model? Or accuracy? Or is it based on some previous study?

 

Response #1: Based on your suggestion, in Table A10 of Appendix A (line [402]), we added the Inception Score and Fréchet Inception Distance corresponding to the composite image when comparing different lambda. At the same time, we also added the calculation of Pearson correlation coefficients of different p at Table A9 and Table A10 (line [407-411]) to verify that our data are not relevant. We are very sorry for the mistakes we have made in the details.

 

Reviewer #2: The same for bi-linear interpolation, what bi-linear? Adding some discussion to justify these choices can help reader to better understand these choices.

 

Response #2: This is a negligence of our work. We have added a lot of explanations to different definitions in the revised paper. Definition of bilinearity as you mentioned (line [216]). At the same time, we have added comparative experiments in every place where experimental comparisons are needed for explanation (line [377-423]). To help readers understand the rationality of each choice as much as possible and describe the relevant nouns. We will pay attention to this issue in future scientific research work and maintain a more rigorous attitude to conduct research work.

 

Reviewer #3: There are still some grammatical error in this paper, and these need to be fixed.

 

Response #3: We are very sorry for the problems caused by the grammatical errors. In the revised version of the paper, we consulted as many professionals as possible to solve grammatical issues, include graduate students in English, MDPI English editors, and other relevant English professionals. I hope to solve the problem of grammatical errors, so that readers can read the paper more easily.

 

Reviewer #4: In Table A.1, it would be great to add the IS and FIS of some of the more recent GAN models too, beside WGAN and DCGAN.

 

Response #4: With reference to your suggestion, we have added a comparison experiment of the latest Self-Supervised GAN [c] and Improved Self-supervised GAN [d] in Table A1 of Appendix A (line [361]). It can be seen that due to the complexity of the lesion structure itself, our IS and FID have not improved significantly. This also confirms the effectiveness of our method. When the direct use of GAN can not generate a better image of diseased leaves, it can first generate a local and then synthesize with healthy leaves to expand the data. The problems encountered when synthesizing images can be solved using edge smoothing or other methods. We believe that this idea can also solve some problems when GAN generates other images.

 

Reviewer #5: A more detailed discussion on the training/test time of this approach would be helpful for readers who want to follow a similar idea.

 

Response #5: We refer to your suggestion and describe the time it takes for us to train our method on the GTX 1070Ti (line [252,307]). At the same time, in order to make it easier for readers to use our method, we submitted the code of our method on Github —— https://github.com/Ronzhen/ES-BGNet (line [252]). We hope that it can help more readers to carry out related scientific research and help more fruit farmers and others to prevent plant diseases.

 

Reviewer #6: Since the main idea of this paper is to use GAN for data synthesis, the authors should add more recent works which use this idea for data augmentation. This idea has been heavily used for image classification, person re-identification, and various biometrics such as fingerprint, iris, and face. The following papers have some of the relevant works:

[a] "Unlabeled samples generated by gan improve the person re-identification baseline in vitro." Proceedings of the IEEE International Conference on Computer Vision. 2017.

 

[b] "Biometric Recognition Using Deep Learning: A Survey". arXiv preprint arXiv:1912.00271. 2019.

 

Response #6: We have increased our number of references based on your suggestions. The previous modification was too focused on the research of GAN in the field of plant generation. In this modification, we have added many other references, such as the latest research on data enhancement by GAN, as you mentioned [a], We added other references (line [99-100,118-128]). At the same time, we have also added references to latest research of deep learning, such as [b] you mentioned, and other references we added (line [79-86]). We hope that our work can be better displayed to readers, and include sufficient references to facilitate readers to retrieve related literature.

 

Thank you again for your valuable suggestions, because of your suggestion, I also gained a lot in this paper work. We will work harder in future research work to improve our ability in scientific research, English, writing and so on.

Have a nice day.

 

[a] "Unlabeled samples generated by gan improve the person re-identification baseline in vitro." Proceedings of the IEEE International Conference on Computer Vision. 2017.

[b] "Biometric Recognition Using Deep Learning: A Survey". arXiv preprint arXiv:1912.00271. 2019.

[c] Self-Supervised GANs via Auxiliary Rotation Loss

[d] An Improved Self-supervised GAN via Adversarial Training

 

Author Response File: Author Response.docx