Hardware-Aware Design of Speed-Up Algorithms for Synthetic Aperture Radar Ship Target Detection Networks

Round 1

Reviewer 1 Report

Summary:

This is a very meaningful study. The timeliness of radar target recognition in hardware has always been a major research challenge. This article proposes an accelerated method for GPU oriented SAR ship target detection task model. Accelerate GPU through methods such as low precision quantification and arithmetic operator fusion. And achieved excellent experimental results on the dataset HRSID. Overall, this manuscript is well written, but there are still some issues that need to be addressed.

 

Summary:

This is a very meaningful study. The timeliness of radar target recognition in hardware has always been a major research challenge. This article proposes an accelerated method for GPU oriented SAR ship target detection task model. Accelerate GPU through methods such as low precision quantification and arithmetic operator fusion. And achieved excellent experimental results on the dataset HRSID. Overall, this manuscript is well written, but there are still some issues that need to be addressed.

 

The paper can be improved in the following aspects:

Detailed comments:

1. When using abbreviations, please provide the full name where it first appears, such as “PTQ”, “QAT”, etc. on page 3, line 108;

2. On page 9, line 360, you said, “There is a conflict among the three optimization objectives mentioned in Equation 13.”, Does “the three optimization objectives” mean in Equation 12 ?

3. In section 4.2.2, the author provides a comparison of the ablation experimental results with TensorRT algorithm. I believe that in section 4.2.1, it is also necessary to provide some figures to more intuitively demonstrate the superiority of the algorithm.

4. In Table 4, some numbers are marked in bold. What does this mean?

5. In Equation 1 and 2, please explain what “round” and “clamp” mean.

Author Response

For details, please check the attached document.

Author Response File: Author Response.docx

Reviewer 2 Report

1. For the first time, the authors introduced the quantization method to the SAR ship target detection network and improved the quantization method according to the problem of the small varieties of targets in SAR ship targets detection. On the basis of quantization speedup, an accuracy adaptive scheduling method is also proposed to make the accuracy loss negligible. The experimental results demonstrate that the SAR target detection network detection speed is accelerated while the accuracy loss is negligible.

2. However, there are still some unclear method descriptions in the article that need to be answered towards improving this article.

3. In Section 3.1.2, when describing the method for calculating the T-value, the thresholds for the corresponding upper and lower limits of the T-value are not described, resulting in a somewhat ambiguous calculation process for the method.

4. In the method of precision-aware scheduling, there are some parameters that are not clearly described in the method for how to find the optimal decision in the Pareto solution set, such as why the RI in line412 goes to 0.58.

5. In line30, why do you need to inverse quantize after quantizing?

6. For quantization, is the choice of 16bit quantization better than 8bit quantization, and what does this acceleration depend on?

7. More references closely related to SAR ship detection can be cited and reviewed, such as: doi: 10.1109/TGRS.2023.3251694, DOI: 10.3390/rs13214209 and , doi: 10.1109/JSTARS.2020.2997081.

8. There are some grammatical errors in the article that need to be corrected. For example, the quote in line360 should be Equation12

1. There are some grammatical errors in the article that need to be corrected. For example, the quote in line360 should be Equation12

Author Response

For details, please check the attached document.

Author Response File: Author Response.docx

Reviewer 3 Report

This article mainly studies synthetic aperture radar ship target detection networks based on hardware-aware design of speed-up algorithms, which has a certain degree of interest. However, I believe that this paper currently does not meet the publishing requirements.The following are the main issues that I am concerned about:

1. At the beginning of the article, the author pointed out that synthetic aperture radar target detection algorithms based on convolutional neural networks have received increasing attention in recent years. However, no detailed comparative analysis was provided between CNN and the newly proposed hardware-aware model speed-up method.

2. The method proposed in this article is based on the most widely used neural network computing hardware - graphics processing units. So what is a similar method for comparison? What role does the CPU play in the experiment?

3. Regarding Figure 5, the name annotation of the abc small image is incorrect. Please make careful modifications. And the yellow and red boxes in the figure are not obvious and need improvement.

4. The author proposes a new model acceleration method suitable for SAR ship target detection algorithms in the article, which uses embedded GPUs as the target deployment platform. However, in the introduction section, there is no detailed explanation of the overall structure of the article, and there is a lack of organization and explanation of the article framework, which is a bad experience for readers.

Minor editing of English language required.

Author Response

For details, please check the attached document.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

It is ok.

It is ok.