Unsupervised Radar Target Detection under Complex Clutter Background Based on Mixture Variational Autoencoder

Round 1

Reviewer 1 Report

The paper develops an algorithm to improve target recognition by addressing the challenges caused by clutters in the radar scanning range without labeling new data. The core idea is based on using a variational autoencoder to model and then estimate the transformed version of the input distribution as a GMM in its bottleneck. As a result, the model becomes robust and stable with respect to distribution shifts. Experimental results are provided to demonstrate that the method is effective and compares favorably against prior works.   The paper tackles a practical valuable problem because tackling distributional shifts in target recognition is crucial. The paper is written well, easy to follow, and the experiments are convincing. I have the following comments to be incorporated before final publication:     1. Simulations are based on using synthetic data. I think this work becomes significantly better if experiments on real-world data are provided.   2. In Figure 13, could provide more explanation why increasing the Z dimension after 5 is not very helpful? If the data becomes more complex, will it affect the observation?   3. From Figure 14, we see that using GMM is a good approximation. What happens if we use more than one Gaussian component for each cluster to approximate the distribution better?   4. The idea of using autoencoders to address distribution shifts using unannotated data, has been used in the AI community before:   a. Affine variational autoencoders. In International Conference on Image Analysis and Recognition (pp. 461-472). Springer, Cham.   b. Detection of Dataset Shifts in Learning-Enabled Cyber-Physical Systems using Variational Autoencoder for Regression. In 2021 4th IEEE International Conference on Industrial Cyber-Physical Systems (ICPS) (pp. 104-111). IEEE.   c. Bi-shifting auto-encoder for unsupervised domain adaptation. In Proceedings of the IEEE international conference on computer vision (pp. 3846-3854).   d. Lifelong domain adaptation via consolidated internal distribution. Advances in Neural Information Processing Systems, 34, pp.11172-11183.   e. SSDAN: Multi-source semi-supervised domain adaptation network for remote sensing scene classification. Remote Sensing, 13(19), p.3861.   I think the above work should be included in the introduction section, particularly because the section does not include other works that use a similar approach.  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

See the attached word file. 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper proposes a radar target detection method based on deep learning in the unsupervised way. The GM-CVAE network model is employed to extract the clutter feature in complex clutter background. Simulation results are provided to verify the superiority of your proposed algorithms. Overall, this paper are both convincing and complete. I have the following suggestions:

1. Abstract: It’s a bit long for this part. It’s better to condense the abstract into a limit of 180~200 words.

2. Many other signal detectors, e.g., information geometric detectors [1], and subspace detectors [2], should be added to enrich this part.

[1] X. Hua, Y. Ono, L. Peng and Y. Xu, "Unsupervised Learning Discriminative MIG Detectors in Nonhomogeneous Clutter," in IEEE Transactions on Communications, vol. 70, no. 6, pp. 4107-4120, June 2022.

[2] Z. Wang, G. Li and H. Chen, "Adaptive Persymmetric Subspace Detectors in the Partially Homogeneous Environment," in IEEE Transactions on Signal Processing, vol. 68, pp. 5178-5187, 2020.

3. Please cite some references for the existence equations.

4. The comparison methods are too old. It’s better to compare your methods with the types of ANMF, subspace detector or other newly algorithms.

5. A major problem often occurs when the deep learning is used for radar target detection, how to maintain the CFAR property. I’m wonder if your proposed method possesses the CFAR property. Please analyze this property by giving the curves of the threshold vs the parameters of clutter characterization, or the curves of Pfa vs the parameters of clutter characterization with a fix threshold.

6. Experiments performed on real radar clutter are also preferred to validate the advantage of your algorithms.

7. Conclusion: It’s better to point out the potential limitations and further improvements of your work.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have addressed most of my comments. It seems that access to real data in this domain is not straightforward.

Reviewer 2 Report

The authors have well answered all my comments. 

Reviewer 3 Report

The authors have addressed all my concerns.