SAR Target Incremental Recognition Based on Hybrid Loss Function and Class-Bias Correction

Round 1

Reviewer 1 Report

In this paper the knowledge distillation method has been incorporated into Synthetic Aperture Radar target incremental recognition with the combination of old sample preservation. In addition, an effective loss function for class separation has been designed, increasing/decreasing inter-class / intra-class sample distance, facilitating thus the separation between new and old classes. Finally, a bias correction layer has been designed and trained which corrects the biased output of the model.

Such as the paper looks interesting. However for the general reader's convenience a few improvements are necessary, notably:

1) A very brief mentioning of the remote sensing aspects must be added. See for example: "New aspects of global climate-dynamics research and remote sensing." International Journal of Remote Sensing 32.3 (2011): 579-600.

2) Give an estimate of the standard deviation of the points shown in Figures 3,5,6 and comment on this, accordingly.

3) Increase the letters size in axes labels of Figure 7

In conclusion, I recommend publication of the revised version of the paper with the consideration of the above-mentioned improvements.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The article deals with the topic of everpresent problem in adaptation of a NN to new dataset challenges, preserving a trained structure with decent resilience to new data classes. The article is well structured and written, results are clearly elaborated and results supports stated thesis.

Author Response

On behalf of my co-authors, I would like to express our great appreciation to your comments concerning our manuscript entitled " SAR Target Incremental Recognition based on Hybrid Loss Function and Class-Bias Correction". Thank you and best regards.

Reviewer 3 Report

First, I would like to congratulate the authors for their excellent work. I did not have the time to carry out a complete review of the paper, and I hope the other reviewers have reviewed it more carefully. The work proposes a new incremental approach by recognizing problems in the use of traditional BIAS. In the tests performed, the accuracy of the developed method is superior to traditional methods. The more examples it stores, the greater the accuracy.

1) I missed dataset images in the article;

2) The results leave me in doubt, could the training have been over-fitted due to the accumulation of data already seen on the network, and why is the accuracy greater?

As I mentioned, the work is exciting and complex, and it would take more time to review it as it deserves to be reviewed; for now, these are my comments.

Author Response

Please see the attachment.

Author Response File: Author Response.docx