MFFA-SARNET: Deep Transferred Multi-Level Feature Fusion Attention Network with Dual Optimized Loss for Small-Sample SAR ATR

Round 1

Reviewer 1 Report

This manuscript focuses on the deep transferred MFFA network for the SAR ATR with dual optimized loss for small samples. There are some issues.
1.The title makes me confused.
2.Even in Abstract, there are many errors. The word "MFFA-SARNET" in line 23 is not abbreviated from the words in line 22. The noun "task" in line 20 is countable, but the noun "task" in line 24 is uncountable. In line 24, the word "MFFA" is not abbreviated from "multi-level feature attention network". the nouns "method" (in line 25), "task" (in line 29), "database" (in line 30) and "condition" (in line 34) need "a" or "s". Two verbs "yield" (in line 31) and "surpass" (in line 32) are not separated by "to".
3.Many nouns in this manuscript need "a", "the", or "s".
4.The combined nouns in lines 89-90 are too long. It is not proper.
5.Eq. (2) is weird. The left side does not have "l", but the right side has "l".
6.How to do the function "f" in Eq. (2)?
7.What is the function "tensor"?
8.The parameter "W" in Eq. (2) has "l", but the parameter "W" in Eq. (3) does not have "l".
9.Why do the "tensor" functions in Eq. (3) have subscripts "o", "p", and "q"?
10.What is the function "A^c" in Eq. (7)?
11.What is the function "A^s" in Eq. (12)?
12.What are the values of the bias terms "b^vc", "b^qc", and "b^c" in line 228?
13.The parameters "V^c" in lines 232, 234, and 236 should be replaced by the parameters "V#c".
14.What are the values of "W#l" in Eq. (2)?
15.What is the parameter "Q" in Eq. (5)?
16. What is the parameter "V" in Eq. (8)?
In conclusion, some abbreviated nouns don't correctly represent corresponding noun phrases and some abbreviated nouns cause confused. Moreover, many parameters need explanations and there are many grammar errors. This manuscript needs major revision.

Author Response

      Dear reviewer, thanks for your efforts, we have uploaded an attachment.

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

In a reviewed paper, the authors proposed a new automatic target recognition method for SAR images which provides improved classification accuracy in the case of small number of samples.

The paper provides a fairly detailed description of the method, the effectiveness is confirmed by numerous computational experiments. The paper deserves to be published, although I have some comments.

The language of the article is generally good, but additional proofreading is highly recommended.

Please check for abbreviations such as CNN-MLP (Line 73), MSTAR (Line 118), SSIM, and others. Despite the fact that many of them are generally recognized, they should be explained at the first appearance in the text.

The designations (fc1-fc2) found on line 155 are explained only after a few pages in Fig. 9.

In equation (1) please indicate what P_i^l is (although later it becomes clear from the text).

The explanation of Table 7 is not clear. If the proposed method is called Net1, then how can it be better than Net1 and Net2? (Line 421). In general, in the paper, information about Net2 and Net3 is provided first, and only after a few paragraphs does Net1 appear.

The same remark applies to Tab. 7, which is explained earlier than Tab. 5 and Tab. 6.

Transfer learning is stated as one of the research objectives, but it is described very briefly in two paragraphs in sections 4.5 and 4.6. I would like to see a more detailed explanation of the used technique and parameters tuning.

Author Response

    Dear reviewer, thanks for your efforts, we have uploaded an attachment.

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This paper provides a machine learning based technique for automatic target recognition in SAR images. Specifically, the authors propose a transferred multi-level feature fusion attention network based on a dual optimized loss in a reduced sample scenario. The approach is validated on the challenging MSTAR dataset also in comparison with other already existing methods.

The paper is overall well-written and organized, even if some sentences are hard to read and understand; moreover, the used Notation should be uniquely defined and carefully checked.

See for instance,

Line 201: “A Suppose that the size of feature map … ”;

Line 270: “If the input of optimization part is xi”;

The symbols in eq. (13) should be modified. Why do you use the arrow in place of =?

In which way one can select the parameter beta? It is not clear to me.

A discussion about complexity should improve the quality of the paper.

In Figure 10, I think that the table is sufficient to show the results, therefore I suggest to remove the histograms.

Some recent papers concerning the problem of ATR in SAR images following a feature-based approach are missed. Please, add a discussion on them in the Introduction:

[1] “Automatic Target Recognition of Military Vehicles with Krawtchouk Moments”, IEEE Trans. on Aerospace and Electronic Systems, 53(1):493-500, February 2017.

[2] “SAR Automatic Target Recognition Based on Dictionary Learning and Joint Dynamic Sparse Representation”, IEEE Geoscience and Remote Sensing Letters, vol. 13, no. 12, pp. 1777-1781, Dec. 2016.

[3] “Man-Made Object Classification in SAR Images Using 2-D Cepstrum”, IEEE Radar Conference, Pasadena, USA, pp. 1-4, May 2009.

[4] “Introduction to Radar Target Recognition”, IEE Radar Series, Institution of Engineering and Technology, 2005.

Author Response

    Dear reviewer, thanks for your efforts, we have uploaded an attachment.

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

This revised manuscript is ready to be published.