CHFNet: Curvature Half-Level Fusion Network for Single-Frame Infrared Small Target Detection

Round 1

Reviewer 1 Report

Dear Editor, In this paper, the novel CHFNet method for IRSTD was proposed and the curvature attention mechanism was developed. There is a detailed analysis on the current research status and relevant methods. It is significant for small target detecting on the IR image with the temperature information. The manuscript can be accepted after the following minor revision. 1．Please provide a detail description on the ‘stitch the resulting eight matrices as eight channels into a new matrix’ in Line 210. 2．It is recommended to adding a noised IR image for better description on the noise attenuation, such as the IR results in Fig.7. 3．In Line 225-227, ‘In this case, except for the higher grayscale values around the two target locations, all areas remain at the lower grayscale values even after 1000 times increase’. The result after the curvature attention branch in Fig.3 is much obvious that the input image. However, it could be necessary to provide a simple example for better understanding on the 1000 times increase. 4．The algorithm results listed in Fig.4 and Fig.5 were not uniform and less than the algorithm number in Line 263-265. Please show the reason to avoid the misunderstanding

Comments for author File: Comments.pdf

Author Response

The authors would like to thank you for all your approval and valuable suggestions. We have considered your suggestions carefully and made every effort to respond to every question raised in the review report.

Author Response File: Author Response.pdf

Reviewer 2 Report

A clear paper devoted to the challenging problem of small targets detection in an infrared image.

The authors provided a good overview of the known identifying features of targets in infrared images, such as size, shape, contrast, etc., but in current study they focused mainly on curvature. How other identifying features affect the classical CNN outputs has not been investigated over noisy infrared images. Although in my opinion, each of the known identifying features is worthy of a separate attention mechanism – this could improve the CNN final result significantly.

As well, there are also pending issues left regarding the curvature engagement, for example – what is the minimum size (in pixels) of the target for which the curvature is applicable? What is the minimum detectable value of curvature and how does it relate to the noise level in image? What types of spatial and frequency filtering can improve or conversely worsen the feature of curvature? I hope that the authors will answer these questions in their future papers.

Unfortunately, the NUAA-SIRST dataset used by the authors contains lot of ground-truth images, relative to which it is impossible to approve that just the target is marked in ones.

There are minor comments on the paper layout:

There is disorder in references within the introduction, so it is impossible to understand which paper the author refers to.

The logics is broken in equation (4): the same operation V₁(z) is designated differently – f₁ and z₁, which makes it difficult to understand. It is recommended to leave a single designation only.

Anyway, the work is deserved to publication in the Remote Sensing journal.

Author Response

The authors would like to thank you for all your approval and valuable suggestions.We have considered your suggestions carefully and made every effort to respond to every question raised in the review report.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Authors,

I have reviewed the "CHFNet: Curvature Half-level Fusion Network for Single-frame Infrared Small Target Detection" paper submitted to the Remote Sensing Journal.

 While the issue of IRSTD is interesting, and the presented solution gives good results, the quality of presentations and used language applied in the submitted article is qualitatively different from the standards of the Remote Sensing Journal. Also, the structure of the study makes the paper very hard to read.

The authors should think about the construction of the whole article. Due to the theoretical sections of the manuscript and their content being introduced by MDPI, authors should use them to allow potential readers to understand what is proposed in the paper easily. Firstly, the introduction should briefly describe the current state of the research related to IRSTD. They should be reviewed carefully, and the description of the publications not only mentions the title of selected papers (mainly used by authors) but the advantages of these studies should be presented. Please remember that scientists working outside the paper's topic might also read the article. Based on that review, it is easier to point out the advantages of the presented method. Then Authors should explain their approach. After that, the test scenario and used parameters for comparison should be given. Finally, the results obtained by the authors (using the presented method and comparison methods) should be clearly shown. After that, authors can discuss the differences, their correlations, etc.

Some other comments are below:

1)The references in the text are not converted to citation style.

2)It would be good to give the text of the paper to a native speaker because there are many omissions and some places are hard to read or look like from google translate.

3) Figure 1 should be placed after it is mentioned in the text.

4) the description of the used parameters should be placed before it is used in Table 1.

5) As far as 13 algorithms (methods) were mentioned in the text, they should appear in the comparison. However, only a few of them were presented in Figures 4 and 5.

 

 

Sincerely,
Reviewer

Author Response

The authors would like to thank you for all your approval and valuable suggestions. We have considered your suggestions carefully and made every effort to respond to every question raised in the review report.

Author Response File: Author Response.pdf