Deep Learning-Based Water Crystal Classification

Round 1

Reviewer 1 Report

Summary

This paper presents a convolutional autoencoder framework to classify the category of water crystal from the proposed new dataset, 5K EPP that contains water crystal images in high resolution.

Overall, the paper seems to be in its initial form with many issues unaddressed (see detailed comments below) that prevent it from being accepted.

Major comments:
1. This manuscript needs careful proofreading. There are typos and issues with grammar and expression throughout the manuscript, which impede understanding. Also, Figure 1 is missing.

2. The motivation of this paper is unclear. Indeed, understanding water crystal helps improve water quality, but what motivates this research and what challenges are addressed remain unclear. More importantly, the manuscript does not provide a clear definition of the term "water crystal" as claimed in the conclusion.

3. The description of the dataset is vague. As the goal of this manuscript is about applying the CNN-based image classification framework to the new dataset, this section is important. The current version of Section 3 is short and islack of details such as the source of the data, specific 13 categories, and labels. Examples would be helpful as well.

4. From Figure 6, the observation is "the training progress can show that our model is more stable and it converges faster". However, this cannot be seen from the figure, as the two curves converge similarly, and the proposed method does not seem to be more stable than the baseline.

5. The novelty of the manuscript is unclear. The components of the proposed framework are widely studied, such as CNN, ResNet, encoder-decoder, etc. It is recommended to have clear statement about the novelty of this work.

 

Minor comments:
1. In line 173, there is an error, "Considering a data sample X with n sample and m features". X seems to be the dataset, rather than an individual sample. Similar errors can also be found in other places throughout the manuscript.
2. The caption of Figure 2 is not very helpful. A more detailed explanation is needed.

Author Response

Dear Expert,

we would like to thank you for your review. We uploaded the answers to your comments and questions.

thank you again for your great help.

All best wishes

Frederic Andres

Author Response File: Author Response.docx

Reviewer 2 Report

Introduction

Due to climate changes invoked by human irresponsibility’s, nature is suffering. Water in the world is more and more less. In this paper and in the aim to contribute to the sustainability of water resources, by the monitoring of water quality, authors present deep-learning based classifier for water crystals. With the Emoto Peace Project, the dataset 5k-EPP is constructed. Some augmentation methods have been applied to the dataset to avoid unbalanced classes. Features are first extracted using a Residual AutoEncoder. Then fine tuning is applied on 5 pre-trained models chosen based on their success in the competition ILSVRC.

Remarks

1- The unique dataset used in this study is 5K-EPP. Unfortunately, there is neither reference to this dataset nor to the laboratory where this dataset was constructed.

2- In Section 3, which is dealing with the dataset presentation, examples of such images are needed like.

3- Also, in Section 3, authors have mentioned a tree-like diagram for the construction of the 13 categories of the data. There is no figure for that ( line 158)

4- Authors claim that 5K-EPP is the first world-wide dataset for water solid crystals. I think that more comprehensive details about the construction of this data will have a valuable added value to the paper.  

5- in line 352, authors claim that the F1-score is much lower but when we return to the table 2, we did not found this.

6- In the experiments, the proposed model is compared with the one in ref [6] on the same dataset 5K-EPP. Results show that the proposed model is more stable and performs better. I have the following suggestion regarding this

• First, the model in [6] was constructed for classifying snowflakes and the dataset is about water crystals. Can this fact affect the performance of the competitor model?
• Using more models for comparison will enhance the results and highlights the proposed model performance.
• Also the is no reference to the dataset MASC (line 364)

7-  In the introduction and in the conclusion, authors claim that they have set 13 categories for the set of images in the dataset. In line 334, it is written that “Finally, we have a new definition with 12 categories, which is defined in 3”. Which is the correct value?

 

2- Some writing errors (among others).

• Line 49: “that provide” instead of “provides”
• Line 53-54: “These features are 54 not only useful for specific but also can help in other related tasks” can be replaced by “These features are not only useful for specific tasks but also can help in other related ones”
• In section 4, authors should number the title “Residual Auto-Encoder” by 4.1.1 and correct “4.1.1 fine-tuning model” by “4.1.2”
• There are many references like “on 3” in page 6 or “in 5” in line 261for example. The reader cannot understand if it is a figure or a section or a table! Authors should clarify to which exactly they are referring.
• Lines 302-303 and 304 are duplicated.

 

Comments for author File: Comments.pdf

Author Response

Dear Expert

 

We would like to thank you for your interesting comments and questions.

gratitude for your great help.

 

Frederic Andres

Author Response File: Author Response.docx

Reviewer 3 Report

The introduction of a new dataset on water crystals and the reported deep neural network based classification could be helpful for other researchers. The classification method is sound and the experimental results support the proposed method. However, there exist some errors requiring correction:

• The paragraph from line 47 to line 57 is redundant.
• The authors should describe the availability of the 5K EPP Dataset more clearly.
• In line 158, “We build a tree-like diagram in Figure ??”.
• Figure 1 is omitted.
• In a line between line 209 and 210, there is a statement of “The final architecture is shown on 3”. It is unclear what “3” denotes in the statement.
• In line 245, “All the experimental results are shown in 5”. It is unclear what “5” denotes in the statement.
• In line 255, “The overview of the final classification model is given in 4”. It is unclear what “4” denotes in the statement.
• In line 261, “The comparison results are then shown in 5”. It is unclear what “5” denotes in the statement.
• In a line between line 276 and line 277, “we set the number of classes equal to the number of ground-truth categories that were used to label the dataset in 3”. It is unclear what “3” denotes in the statement.
• In line 303, “The reconstructed images generated by the model trained with BCE and Spherical are shown respectively in 5”. It is unclear what “5” denotes in the statement.
• The statement in line 302, “The reconstruct results built with both metrics are shown in 5”, is redundant.
• In line 313, “We train the classification models proposed in 4”. It is unclear what “4” denotes in the statement.
• In line 334, “Finally, we have a new definition with 12 categories, which is defined in 3”. It is unclear what “3” denotes in the statement.
• In line 334, the authors stated that they introduced a new definition with 12 categories. But it is not clear what the new definition is. The authors did not contrast the old definition with the new definition.
• In line 340, “We fined-tuned parameters and applied data transformation techniques mentioned in 3 to enrich the dataset”. It is unclear what “3” denotes in the statement. In addition, the authors did not report the data enhancement or data enrichment result.
• In the “experimental results” section, some quantities are more suitable with “percentage point” than “percentage”.

Author Response

Dear Expert,

we would like to thank you for your review. We uploaded the answers to your comments and questions.

thank you again for your great help.

All best wishes

Frederic Andres

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The reviewer would like to thank the authors' attempt to revise the manuscript. However, because the reviewer's comments in the previous round are not sufficiently addressed (see details below), it is not sufficient to consider the acceptance of the manuscript in its current form.

Details.

Major comment 1:
The manuscript in its current form still has many issues wrt grammar/expression, which includes but are not limited to the following:
    1. Line 59, "now are used widely" -> "are used widely", "includes" -> "including".
    2. Line 385, the reference to the figure is erroneous. It should be section 5.3.3, or Figure 6.
The reviewer can only point out several examples when reading the updates. Please consider a serious round of proofreading as well as grammar/expression checking.

Major comment 2:
The new table 3 is helpful. In the main text, it is recommended to add a general definition or a short description of "water crystals" and then refer to Table 3 for more details.

Major comment 4:
In Figure 6, the performance of the proposed model (the green curve) stops to improve at iteration ~ 65, and this is the same as the baseline (the blue curve). Therefore it is unwarranted to claim that the proposed method converges "faster". Indeed, the proposed method seems to perform better overall throughout the training process, but the speed of convergence is not shown in the figure.

Major comment 5:
Please consider adding the novelty you claimed in the main text.

Minor comments 2: The caption of Figure 2 does not seem to be updated. It is recommended to explain in the caption of figure 2 about the explanation why the model is proposed, and what effect it plays compared to the original residual block.

 

Author Response

dear Expert

 

we would like to thank you for your review.

 

All best wishes

 

Frederic

Author Response File: Author Response.docx

Round 3

Reviewer 1 Report

The authors have attempted to address most of my previous comments. I would suggest accepting the manuscript after the manuscript has gone through a round of proofreading and grammar checking.