End-to-End Powerline Detection Based on Images from UAVs

Round 1

Reviewer 1 Report

Overall, the research seems to be practical and applicable in the real world. The methods used are sound and appropriate and demonstrate practicality. The authors have well stated the need for this research and the application of UAV in the research. However, the citations, grammar, and structure of the paper needs to be improved. It is suggested not to use active phrases like “we” in the paper. And it is suggested that the authors follow the template of the journal which is not met in this research.

 Page 1, Line 21, Abstract, f1score, should it be f-score?

Page 2, Line 49, 1. Introduction, What are the traditional methods?

Page 2, Line 50, 1. Introduction, Mis detect and Misdetection, the same words are used twice.

Page 2, Line 50-52, 1. Introduction, Please add a reference to this line.

Page 2, Line 55, 1. Introduction, Please abbreviate the full form of RCNN for the first time you use it.

Page 2, Line 62, 1. Introduction, Please abbreviate the full form of FPN for the first time you use it.

Page 2-3, Line 73-145, 1. Introduction, It might be better to disintegrate this into a separate section as Literature Review.

Page 2, Line 96, 1. Introduction, Given its unstable performance, how is this disadvantage of Hough Transform tackled in the study?

Page 3, Line 105, 1. Introduction, Please abbreviate the full form of YOLO for the first time you use it.

Page 4, 2. Methods, 2.2. Hough FPN, Figure 2, Please name the axes.

Page 5, Line 186-189, 2. Methods, 2.2. Hough FPN, what will be the effect of transforming the images into these resolutions? Have they been considered in the study?

Page 6, Line 193, 2. Methods, 2.2. Hough FPN, Is there a specific reason to use 2-D instead of 3-D?

Page 9, Line 278, 3. Results, 3.1. Dataset, How were these images captured? The capture parameters should be reported like the flight height, flight angle, scale of the image.

Page 9, Line 294-295, 3. Results, 3.2. Implement Details, 3.2.1. Experiment Environment, The average inference speed should be reported. Why was the average inference speed calculated for only 50 images instead of 1497 images from the dataset?

Page 10, Line 299, 3. Results, 3.2. Implement Details, 3.2.2. Training Detail, SGD should be abbreviated for the first time it is used.

Page 10, Line 330, 3. Results, 3.3. Metric, The effects/errors caused by scaling should be accounted in this study.

Page 11, 3. Results, 3.4. Comparison with other methods, Table 1, Is the precision that is obtained by your method after considering the errors? Like some errors might have arisen due to pixel’s effect in the image, assumption of straight network line and scaling of the images.

Page 11, 3. Results, 3.4. Comparison with other methods, Table 1, Overall these values are only valid if the transmission network line is perfectly straight. Is there a way to account for the errors caused due to the curvature of the network lines?

Page 14, 3. Results, 3.5. Ablation Study, Figure 9, The axes should be named.

Page 14, Line 417-420, 3. Results, 3.5. Ablation Study, Any reference on this statement?

Page 16, 4. Discussion, A brief comment on the ways to tackle high amount of obstacle in the transmission line detection should be included. 

Author Response

Dear Reviewer,

Thank you for reviewing our manuscript titled “End-to-End powerline detection based on images from UAV” and providing valuable feedback. We apologize for any inconvenience caused by our negligence in preparing the manuscript. We have carefully revised the manuscript based on your comments and attached  detailed response to each of your comments in a separate document.

Thank you for your time and effort in reviewing our work. We hope you find the revised manuscript satisfactory and look forward to your further comments.

Sincerely

Author Response File: Author Response.docx

Reviewer 2 Report

Summary

This paper proposes a method for the detection of transmission lines in images based on the use of a novel neural network architecture.

Review

The article contains one interesting idea, namely, the incorporation of the Hough transform in the architecture of the network. Is it original of the current paper or can be found in the existing literature?

However the article cannot be recommended for publication in its current state since it presents several flaws:

- The English writing must be completely revised: some expressions are hard to understand, there are several typos and syntax errors.

- The assumptions on the characteristics of the transmission lines (straight lines) do not always hold in a general situation, as the authors themselves admit in the final Discussion of the paper. This makes me doubt the applicability of the proposed method in practical situations.

- The style of description of the previous literature needs to be improved: replace sentences of the type "[11]Using the corner radius..." by "Duda and Hart [11] use the corner radius ...".

- The proposed architecture must be described in detail. Figure 1 doesn't provide any useful information about number of layers, kernel sizes, types of operations, etc.

- The use of the Hough transforn within the network must be better explained. The sentence "The FPN layer is composed of 186

three characteristic graphs X1, X2, X3 which is respectively 1 / 2,1 / 4,1 / 8 of the resolution of the input image, and then the convolution operation on the Hough domain will be carried out" is too cryptic. More details, and maybe a illustrative figure, are needed.

- The authors claim that they use a "neck structure to provide multi-resolution output". This structure should be described in the text and an illustrative figure shuold be provided.

- In the results section, what does it mean the sentence "the ground truth of picture 512 is scaled to 128" (line 329)?

- In Table 1, what are the values in the Inputs column? Number of images used in the experiment?

- Columns Precision and Recall in Table 1 seem to have switched their names. In order for the results to be coherent with the ones displayed in Figure 5, the column named "Precision" seems to correspond to Recall values, and vice-versa.

- In Figure 6, why short edges are not detected as segments by the proposed network? Is it due to the training?

- The ablation study is just based on the results for two images. The authors should display precision-recall curves such as the one in Fig. 7 for the different versions of the method: with Hough and NMS, without Hough, without NMS

- Minor details:

    - There are 2 figures labeled as Figure 3. The one referred to in line 185 depicting the "FPN structure using the Hough Transform" seems to be missing

    - The paragraph in lines 42-46 in the introduction seems to misplaced

Author Response

Dear Reviewer,

Thank you for reviewing our manuscript titled “End-to-End powerline detection based on images from UAV” and providing valuable feedback. We apologize for any inconvenience caused by our negligence in preparing the manuscript. We have carefully revised the manuscript based on your comments and attached  detailed response to each of your comments in a separate document.

Thank you for your time and effort in reviewing our work. We hope you find the revised manuscript satisfactory and look forward to your further comments.

Sincerely

Author Response File: Author Response.docx

Reviewer 3 Report

This paper describes the development of a method for the detection of transmission lines. Compared to state-of-the-art methods, the developed method performs significantly better. 

In principle, the results are convincing. Personally, I cannot sufficiently assess the new value or impact. If the editor believes that the novelty is given, I am in principle in favor of publication.

 At the moment, however, the form is not yet in a publishable state.

Some sections are difficult to understand and some wording is grammatically incorrect. The equations are incorrectly formatted in the version I have (missing spaces, text too large). Some figures could be optimized and various equation characters are not introduced.

Below are some comments, which are however not complete.

-          Abstract, line 21: f1socre -> F-score ?

-          Line 43/44: It’s not clear which source belongs to what sentence, spaces missing?

-          Line 54: What is the “:” for? The sentence seems not complete;

-          Line 53: yolo small (later capital letters)

-          Please explain aberrations when first used (e.g. YOLO - You Only Look Once, RCNN, PANet, FPS, DWP…)

-          Line 80: trans-form (delete -)

-          Line 138: rcnn not in capital letters

-          Please explain Figure 1 in more detail/ add a description in the figure

-          Font size/spacing of all equations is off,

-          Not all symbols in equation 6 are explained (e.g. what is n, gamma, L?)

-          grad-operator should not be cursive

-          word-indices should not be cursive (e.g. gt_center)

-          eq. 7 : “negitive”

-          line 233 : smooth(?) should probably not be here

-          eq. 7 : N_all, N_positive, N_negative are not explained

-          Lines 248-249: I don’t understand this paragraph. What exactly is the problem with transmission lines and what exactly does NMS fix? Please elaborate.

-          Line 255: what do you mean by “the model detects one single line for multi-times”? From the figure it’s not clear if there are multiple transmission lines or just one, is the same line detected multiple times?  

-          Eq:10, Rec is not explicitly explained

-          Table 1: FPS is not explained (Frames per second? Wouldn’t a higher number be better and your method, therefore, be better than LSD and HoughP? “Although our method may not be comparable to LSD … in terms of FPS” sounds like it's worse.)

-          Figure 5: Font is too small

-          Figure 6b) lines are hard to see, looks a bit like no lines are detected

-          Figure 9: axis missing – what is being displayed here? 

Author Response

Dear Reviewer,

Thank you for reviewing our manuscript titled “End-to-End powerline detection based on images from UAV” and providing valuable feedback. We apologize for any inconvenience caused by our negligence in preparing the manuscript. We have carefully revised the manuscript based on your comments and attached  detailed response to each of your comments in a separate document.

Thank you for your time and effort in reviewing our work. We hope you find the revised manuscript satisfactory and look forward to your further comments.

Sincerely

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

There is a notable improvement in the quality of the paper with respect to the previous version.

Some small typos remain:

- N_negitive in Eq. (9) (N_negative)

- the green curve in Fig. 13 should be "With NMS without FPN", I think

And the English should be further improved prior to publication. For instance, this sentence in the Discussion  is grammatically incorrect (and incomprehensible):

"for the first, analyze the morphological characteristics of the lines in the images, analyze them for the characteristics,"

Author Response

Dear Reviewer,

Thank you for your time and effort in reviewing our paper and your positive feedback. We appreciate your valuable comments and suggestions that helped us improve the quality of our paper. We have carefully addressed all your concerns and revised our paper accordingly. The revised portions are marked in red in the paper. Please see the attachment for the corrections in the paper and our responses.

We hope that these revisions meet your approval and that our paper is now suitable for publication.

Thank you again for your constructive feedback and support.

Sincerely,

The authors

Author Response File: Author Response.docx

Reviewer 3 Report

The authors have largely implemented my comments and therefore I am in favor of publication.

However, the paper should be briefly reviewed again, for example there are still a few missing (or additional) spaces e.g. line 203, 206, 245

In some parts it still says "negative" (should be "negative") 

Line 299-314: should this really be formatted bold?

Line 301: l1 not formatted correctly

Author Response

Dear Reviewer,

Thank you for your time and effort in reviewing our paper and your favorable opinion on publication. We are glad that you found our paper improved after implementing your comments. We have carefully checked and revised our paper again according to your suggestions. The revised portions are marked in red in the paper. Please see the attachment for the main corrections in the paper and our responses to your comments.

We hope that these revisions meet your approval and that our paper is now ready for publication.

Thank you again for your constructive feedback and support.

Sincerely,

The authors

Author Response File: Author Response.docx