An Underwater Side-Scan Sonar Transfer Recognition Method Based on Crossed Point-to-Point Second-Order Self-Attention Mechanism
, Haisen Li 1,2,3,*, Chao Dong 4,5, Jing Wang 6, Bing Zheng 4,5 and Tianyao Xing 1,2,3Round 1
Reviewer 1 Report
The paper proposes a deep-learning method for sonar image recognition. However, I can’t find any special features that are particularly designed for SSS. In other words, they only proposed an ordinary method for ordinary images, which is not creative. The sonar images proposed are just considered as grayscale images and the objects in the images are quite clear, which is easy to recognize with common deep learning methods. Transfer learning is a common technique in deep learning. The proposed PPSHA and PPSVA seem new to me. They are modified from the self-attention technique. But their advantages compared to the common self-attention method are not given. And the motivation to do this kind of modification is not clear. The results in such an easy dataset are not convictive.
The writing skills still need to be improved. Some sentences are difficult to understand.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
To improve the precision of categorizing underwater sonar images under the condition of a limited number of mode types and data samples, this study proposes a crossed point-to-point second-order self-attention (PPCSSA) method based on the double-mode sample transfer recognition. In general, the topic in this paper is very interesting. This paper is well organized and the work is comprehensively described. However, the reviewer has some improvements which should be addressed before acceptance.
1. In Eq. (8), the cross-entropy loss is presented. The reviewer wanders to know how to determine M and N when it comes to the processing.
2. The authors should clearly clarify the difference between their method and traditional method in a single section. With this operation, the readers can easily understand the advantages of their method.
3. In practice, the synthetic aperture sonar [1][2] can also produce sidescan image, which is characterized by high-resolution. The image of traditional sidescan sonar is highly dependent on the array aperture, and it is characterized by low-resolution. The reviewer wanders to know whether the authors’ method can get the same conclusions or results from SAS image and traditional sidescan sonar image. The references should be enhanced. Furthermore, the authors should discuss the performances based on both sonar images, individually.
[1] Zhang, X., Yang, P. and Zhou, M.. Multireceiver SAS imagery with generalized PCA. IEEE Geosci. Remote Sens. Lett. 2023, 20：1502205．
[2] P. Huang and P. Yang, "Synthetic aperture imagery for high-resolution imaging sonar," Frontiers in Marine Science, vol. 9, p. 1049761, 2022.
4. From Table 5, the improvements of authors’ method is observable. The reason behind this should be further enhanced.
5. The experiments in section 4 should be enhanced. On the one hand, the authors’ method should be performed based on SAS image. On the other hand, the authors’ method should be performed based on traditional sidescan sonar image.
6. The English should be improved.
Moderate editing of English language required
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
More inputs in introduction and results are expected. Results section need to be strengthened with more results.
Minor editing is expected.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
1 As I have mentioned in the previous review, the proposed dataset is easy to recognize, and there is no significant difference in the results as the quantity of images is quite limited. Perhaps you can provide some specific images that are recognized by your methods but are not recognized by others.
2 Anyway, the result in FLS dataset is quite remarkable. But the ablation experiments should be provided.
3 And I also suggest that you provide some specific images.
4 In addition, I think Fig. 7 is not required as the image transformation is a common technique.
5 It is better to list processing time and provide your dataset online if possible.
6 Is your method specifically designed for SSS images? Can they be applied to ordinary images? I have this question because the advantage you described seems common to all images.
In Table 1， The model structureis...
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The paper is improved after revision
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
