Ship Classification in SAR Imagery by Shallow CNN Pre-Trained on Task-Specific Dataset with Feature Refinement

Round 1

Reviewer 1 Report

This paper presents a lightweight network with feature refinement pre-trained on task specific dataset for ship classification in SAR images. Generally speaking, the method lacks novelty. Three aspects of improvement are reported, which are pretraining using task-specific data set, lightweight CNN and feature refinement. However, pretraining using task-specific data set has been widely used already, while the lightweight CNN was selected from the 28 models, which have traditional structures, and at present, there are already many advanced lightweight CNN structures. For the last one, the feature refinement is implemented by selecting the convolutional kernels based on the feature map response. This point is not described clearly, and the rationality cannot be judged. The detailed comments are as follows.

1.     Equation 6 is not clear. If the size of the feature map in the last layer is W*H*C, why only compute 2*2 feature matrix? And which position is this 2*2 feature matrix selected from the W*H*C feature map? Is nc the number of training samples? According to the description, 50 kernels with largest COV value for each class are selected. Are the 50 kernels the same for each class?

2.     In algorithm 1, line11: what does this equation mean?

 Line13: do you mean that the features of other kernels not selected are just removed？

3.     In the experiments, the authors only compare their own methods. And the proposed method is not compared with other existing representative methods, such as those for SAR ship classification or representative lightweight CNNs.

Author Response

Please kindly refer to the attached response file.

Author Response File: Author Response.pdf

Reviewer 2 Report

 

The main work of this study is that the number of convolutional layers, the size of the convolution kernel and the pooling method of VGGNet have been modified, and the innovation points are insufficient. In addition, the manuscript has the following problems:

1、In the introduction, the author should specifically introduce the work done by other researchers to solve the problem of insufficient SAR data rather than simply classifying the work done by researchers into two categories.

2、The title of the manuscript is Lightweight networks but neither the introduction nor the abstract by the authors mentions how to achieve lightweight networks.

3、The L2 mentioned by the authors on page 4 should belong to a way of regularization rather than a pooling operation.

4、The experimental part of Section 3 should include the experimental content while the author only introduces the data set and experimental variables.

5、Due to too much data in Table 2, it is suggested that the author highlight the best indicators.

6、The authors did not mention the number of images included in SD1 and SD2 in the dataset introduction, so it is not known whether the authors conducted experiments on large-scale datasets.

7、There is no real discussion. A proper discussion should summarize the main findings and identify the main weaknesses of the proposed method/system, which in turn should be used to propose some directions for future research.

8、The main work of this study is to modify the number of convolution layers, the size of the convolution kernel and the pooling method of VGGNet. It is recommended that authors design some modules themselves to add to the network.

 

 

 

 

Author Response

Please kindly refer to the attached response file.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this manuscript, the authors present a lightweight network for ship classification in SAR images. The network is pre-trained on a task-specific dataset-optical remote sensing ship dataset. Although the experiment results show the effectiveness of the design method, some problems should be considered carefully in the revision. The comments are listed as follows.

1. 2.1. CNN Architecture Design

This part is well known, which suggested to be introduced briefly.

2. The explanation of formulas

Each letter in the formula should be given a specific explanation, such as p and q in Eq. (1), etc. Please check the explanation of all formulas in the manuscript.

3. Experiments

3.1. Dataset and Data Pre-processing

In paragraphs 1 and 2, although references are cited, a brief description of the ORS, SD1, and SD2 data should be given, including how many images there are in total and how many images are in each category.

3.2. Experimental Protocol

Do the "traditional method" in E1, E2, and E3 mean the same thing? I suggest to make a distinguish.

4. Writing

The writing of this manuscript should be improved. For example, Introduction Paragraph 2 Line 2-3, the sentence is too long, please break it into short sentences, etc. Also, please make a comprehensive review of the use of the English language.

Author Response

Please kindly refer to the attached response file.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors did not take our eighth recommendation seriously. Just as I have only verified three hypotheses with VGGNet and its variants, it seems to me that I should make further improvements to the ship detection model based on these three hypotheses.

Author Response

Please kindly see the attached reply.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors have answered all the questions and this paper can be accepted now.

Author Response

Please kindly see the attached reply.

Author Response File: Author Response.pdf