kumoh national institute of technology
Networked Systems Lab.

Thien Huynh-The, Doan Van Sang, Kim, Dong-Seong, "Deep Learning-based Waveform Recognition for Intelligent Radar Systems" , IEEE Wireless Communications Magazine (in Revision), 2020 (IF:11.4)
By :
Date : 2020-09-28
Views : 99

The review of the above manuscript submitted to IEEE Wireless Communications has been completed. The reviewers have recommended that major revisions be made to the manuscript. An acceptance decision will not be made until these revisions have been made and a second review has been completed. Please include detailed responses to the reviewers' comments.

The reviewers' comments are included below. **Please note that there might be files attached to the reviewer's comments. You may not be able to view attached files from this e-mail. To view attachments, please log into your Author Center and click on "View Letter" in the "Processing Status Column" of the table at the bottom of the page.

Once you view the comments, you can respond to the editor and reviewers by clicking on "View Comments/Respond and entering responses in the appropriate fields. To upload a revision of your article, click on its title and follow the instructions. Be sure to complete all pages and click "continue" at the bottom of each one until the revised manuscript has been uploaded successfully.

Once the article has been submitted, you will receive an email stating that fact. If you do not receive an email to that effect, return to the author center at Manuscript Central and check to see if the article is listed in the Partially Submitted queue. If it is in this queue, you should complete the submission at that time. If you are unsure of how to proceed or you do not see your manuscript in the list of submitted or partially submitted articles, contact the administrator (yqian2@unl.edu).

IMPORTANT: It is required that a revision and its revision summary MUST be uploaded to the manuscript central within 45 DAYS after this decision has been made. Failure to do so may cause excessive delay to process your revised paper.

Thank you for submitting your manuscript to IEEE Wireless Communications.

Comments to the Author:
Please prepare a revision based on reviewers' comments.

Reviewer: 1

Comments to the Author
There are minor grammatical errors throughout the paper. I recommend that the authors ask a native-English speaker to proofread and correct those errors.

The "research opportunities" section (VI) should be reduced and incorporated into the Introduction. Placing this section at the end of the paper interrupts the flow and distracts the reader from the results & conclusions of the present work.

Otherwise I believe that the paper is technically sound and sufficiently novel. Good work.

Reviewer: 2

Comments to the Author
In Fig. 1 where deep learning is promisingly discovered for radio signal processing at the device and network levels. Deep learning methods are widely used for signal processing not only at device and network levels, but also at the application level for tasks in Fig. 1, such as emotion recognition, object detection, speech recognition, and etc..

The idea of CNN model with a branch is similar with identity mapping in ResNet model (for one branch) and DensNet model (for multi-branch). Whats the different or special about AMR-Net.

In Section , For example, compared with a regular CNN which adopts 3 x 3 kernels, our network gains approximately 44 percent and 67 percent of fewer parameters with 1 x 5 and 1 x 3 kernels, respectively.. The calculation of network complexity seems not correct which using the size of a single convolution kernel.
For one thing, the 44 percent and 67 percent is can only reflect changes in the number of parameters of the convolution kernel which is not the whole network complexity.
For another thing, in fact, the effect of one 5x5 kernel is equal to that of two 3 x 3 kernels in deep learning, where both them can convert a 15 by 15 image into a 6 by 6 image with no padding and step = 1. So, simply comparing one 3 x 3 kernel and one 1 x 5 or 1 x 3 kernel is not rigorous.

In order to further demonstrate the superiority of the network, more classic DL model should be consider in Method comparison, such as VGG, ResNet, DensNet and etc. Besides, the comparsion should be conducted in some open datasets.

Reviewer: 3

Comments to the Author
Authors consider the problem of automatic waveform recognition, and aim to bypass feature engineering based classification techniques via deep learning. To this end an architecture is proposed that directly processes the received I/Q samples, which is tested on classification of 4 principle pulsed radar waveforms: rectangular, linear frequency modulation (FM), Barker phase-coded, and stepped FM.

The paper has considerable tutorial content on the state-of-the art waveform recognition. Overview of CNN's lack a diagram or a supporting figure to accompany the descriptions provided in Section 3. Authors provide sufficient discussion on their numerical findings, which study the performance of their architecture with respect to the hyper-parameters of learning. The paper is well organized and well written. My comments on the manuscript are as follows:

In Section 4A, authors when forming their training/test sets specify sampling their parameters through the uniform distribution. It is confusing that at a fixed sampling rate, where sampling N (number of signal samples) determines the pulse-width, the number of cycles and the central frequency of the signal would also be uniformly sampled. Yet these are altogether stated as parameters that were sampled from the uniform distribution for all waveforms under consideration. Authors should elaborate on the nature of the training set, perhaps by providing a table of parameters for each waveform class.

Authors among their contributions mention less computational expense than the baseline CNN methods. To this end, it would be beneficial to provide the total training-time and test-time with respect to the baseline deep models in comparison. Furthermore, authors stress the significance of their global averaging and their branch structure within their model, however do not test the effectiveness of this component in their numerical assessments. An effective way to address this would be to modify the concatenation step by leaving out certain global averaging branches during training to compare with the current fully branched model.

Finally, authors for their numerical evaluations only consider 4 classes of waveforms, which is less than those encountered in the literature [*], where several polyphase codes are also considered (Frank, P1-P4). To this end, authors could assess the effectiveness of their model with a wider-range of pulsed radar waveforms.

[*] Zhang, Ming, Lutao Liu, and Ming Diao. "LPI radar waveform recognition based on time-frequency distribution." Sensors 16.10 (2016): 1682.