Rescaling-Assisted Super-Resolution for Medium-Low Resolution Remote Sensing Ship Detection

Round 1

Reviewer 1 Report

This technical note emphasizes adopting SR results as the basic data for object detection, and its quantitative and qualitative experiment results of SR and object detection are impressive to me. In general, this work is complete and with high reference significance. In particular, some novel schemes should be encouraged, for example, comparing and merging rescaling and SR, making the L1 distance between pseudo LR image and adapted LR image. However, there are some problems to be pointed out in the content and legend expression, which I hope the author to improve.

1. As the training process is elaborated in the Experiment, it is hoped to add some presentation elements to show the first and second stages of training in Figure 1. It will facilitate readers to better understand the training process and related losses.
2. As for (a) in Figure 2, pixel shuffle was highlighted. But the article did not introduce how it was implemented. Please expand it
3. As for (c) in Figure 2, the meaning of the four convolutional layers in front of residual blocks is "generate initial feature F₀". However, why four convolutional layers were used, and how to set the parameters of the four convolutional layers? Can the importance of the convolutional layers be proved by experiments? Similarly, please explain the convolution layers after residual blocks.
4. In the experiment, it is mentioned that "Bicubically down-sampled versions were used as the M-LR images". Is it reasonable to replace real low-resolution images with the down-sampled results? The feasibility of the experiment data should be briefly discussed in this article.
5. Line 192 "The image rescaling was trained with the L1 Loss between the rescaled images and original HR images.", it is suggested to integrate the descriptions of the first training stage into 3.3 Loss Function and introduce losses by stage.
6. Line 221, it is mentioned that the ship object should be categorized by scale, but there is no subsequent analysis about the multi-scale ships. Please quantitatively analyze the performance differences of objects with different scales in different object detection algorithms such as RASR, SR, and so on.
7. Line 257, the amplitude-frequency graphs are used for image information analysis. Please describe which method is used to obtain the frequency images in your paper.

Author Response

We thank the reviewer for the constructive comments. In response, we have made a number of revisions to the original manuscript. We believe the paper has undoubtedly benefited from these revisions. The attached file is our detailed response to the reviewer’s comments.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors propose the image rescaling assisted super resolution (SR) method (named RASR) to superresolve remote sensing images. 

The authors provide a overview of evolving trends in the field, but the fact is that there are several unmentioned papers that use some variant of deep learning algorithms.

All the steps of the experiment are recounted in detail, but mainly in the part related to the proposed RASR method. However, the part of the manuscript related to the planning and implementation of the ship detection experiment lacks details. For example, it is unclear how the mentioned training and validation datasets were used? It is necessary to make a clear distinction between test and validation datasets, in the way these datasets are used in a deep learning context (validation dataset: a set of examples used to tune the parameters of a classifier; test dataset: a set of examples used only to assess the performance of a fully-specified classifier). In the same context, how hyperparameters are selected, while avoiding overfitting? 

Why Faster-RCNN and ResNet50 as backbone were selected? What other algorithms have been considered for comparison?

Placement of tables and figures is not optimal. Why Figure 6 is placed at the end of the document?

There are many different abbreviations, their definition in the appendix should be considered in order to increase the readability of the paper.

Author Response

We thank the reviewer for the constructive comments. In response, we have made a number of revisions to the original manuscript. We believe the paper has undoubtedly benefited from these revisions. Below is our detailed response to the reviewer’s comments.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper, a method for image super-resolution for ship detection is proposed.

The paper is hard to read and understand, especially in sections 1-3, due to the adopted cryptic linguistic style, the numerous lexical redundancies, and in particular the lack of clarity. Therefore, a thorough rewrite is required:

• the introduction is not absolutely clear and adequate. The context, the objectives, the novelty, and the modalities should necessarily be better described;
• Section 2 requires further study as the vast literature in the field is only superficially mentioned and lacks a reasoned assessment of existing methods that highlights what is missing and what the paper's contribution is;
• Section 3 is unclear as it is not clear what objectives and solutions are adopted by the authors. For example, more details about the adaption module and its functionality should be given. Also, it seems that the method can be used only one already has a ground truth. If this is the case, the method is of little utility; if not the case, the authors should better explain the method to highlight how it works in the absence of ground truth.

The experimental validation (section 4) is more precise and convincing. However, some aspects necessarily need to be revised:

-For the HRSC2016 dataset, the images are resized to 800x512. How was the resizing done? Does this affect the performance of the proposed method?

- In the experiments, the images of each dataset were used as HR ground truth, and their LR version was obtained with bicubic. But what happens if the LR image is generated by a different downscaling method? This necessarily needs to be studied

- in addition to bicubic other more recent methods not based on deep learning should be considered for the numerical comparison

-  Processing time should be evaluated and compared

- the ship detection should be evaluated also quantitatively

Finally, in the conclusions or another section, the limitations and cases of failure of the proposed method should be highlighted and discussed.

Author Response

We thank the reviewer for the constructive comments.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors substantially improved their manuscript by adding details.

Author Response

Thanks for the comments.