A CNN-Based Method for Counting Grains within a Panicle

Round 1

Reviewer 1 Report

The paper describes a CNN-based method for the grain counting within a panicle by using a CNN. The paper is well written and the analysis carried out well described.

The following are my major comments:

- The algorithm is trained using a sample generated from a single original image. I suggest generating the sample from more original images in order to include more variability in the training set. 

- In the sample generation process, in addition to rotation flip and scaling, I suggest considering also the graining operation.

- The major issue of this work concerns the evaluation process. In fact, the authors evaluate the algorithm using a subset obtained from the same source of the training subset. Consequently, the same original image has been used to generate images for both training and testing. In order to test the algorithm with images totally independent from those used in the training phase, my suggestion is to generate the testing sample from different original images than those used to generate the training sample.

- Both computational time and performance comparison between the CNN-based algorithm proposed and more classical ones (such as image segmentation based on OTSU) on the same datasets would be useful for the readers

Minor comments:
- It could be interesting to add the computational time of the different settings/analyses.
- In equation (3) to replace the words "sum" with a mathematic notation
- To define the acronyms MAE, MSE, MRE as well as to add the reasons the authors decided to use them 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper presents a method for counting grains in a panicle using CNN algorithms. As asserted by authors, the algorithm seems to provide good results for the type of CNN parameters approached.

The following observations and queries may be, however, addressed.

- What are MAE, MSE, MRE? Please explain all acronyms and quantities involved in the text.

- It is not clear how the training set is built. Why artificial enhanced images were employed, since acquiring real images is easy enough to perform.

- What was the success rate for the 4 networks involved? What is the most convenient one? What conclusions may be drawn with this respect?

- Can the method be implemented in the field or only in the laboratory?

- English must be substantially reviewed. Many phrase topics are wrong. The word “and” is often used incorrectly or misplaced at the beginning of the sentences.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The new version of the paper has improved. The following are a few further suggestions:
- the computational time has to be removed from table 2 and add in table 4, for all six methods used (both CNN-based and classical ones).
- The authors should indicate in the manuscript that the diversity of panicle phenotypes is guaranteed by the use of five variants, such as Oryza sativa ssp. Indica variety JP69 and Oryza sativa ssp. japonica variety Jiaoyuan5A.
- the slight improvement of the performance of model 4 with respect to the other three models and to the Wavelet Model gives reason for its elevated complexity and its higher computational time? The authors have to add a discussion about that.

Author Response

Thanks for the reviewer's careful work.Please see the attachment for the details of the authors' responses.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper answered satisfactory to my questions and observations. In my opinion, the paper may be accepted.

Author Response

Thanks for the careful work of the reviewer.