Target Detection in Sea Clutter Background via Deep Multi-Domain Feature Fusion
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis manuscript presents a novel multi-domain feature fusion network (MFN) for small target detection under sea clutter, integrating features from time, frequency, fractal, and polarization domains. The proposed approach leverages an autoencoder-based feature compressor and introduces a false alarm rate (FAR) control mechanism to enhance the practical reliability of detection. Experimental results on the well-known IPIX dataset demonstrate promising detection performance under varying false alarm constraints. The topic is timely and relevant to maritime radar applications, and the method shows a clear engineering focus on robustness and deployment feasibility.
However, there are still several important issues that need to be addressed before the paper can be considered for publication. These include:
- Method Naming and Distinctiveness
The proposed method is named “MFN,” which is quite generic. To improve clarity and citation potential, consider using a more specific name (e.g., incorporating the concept of “false alarm control” or “polarization-aware fusion”).
- No Ablation Studies
While comparisons to baseline methods are provided, the paper lacks ablation studies to assess the contribution of each component. For example:
- What is the performance without the autoencoder?
- What if only three (or one) domains are used?
- What is the impact of removing the false alarm constraint?
This would help justify the inclusion of each module.
- Insufficient formal definition of the false alarm control mechanism
The paper mentions a “false alarm constraint mechanism,” but lacks a clear mathematical formulation or algorithmic description. Please provide a more rigorous explanation and how it is incorporated into the training or inference stage.
- Mathematical Clarity and Notation Cleanup
There are minor issues with notation clarity:
- Some functions (e.g., F for features and for loss) are overloaded;
- The loss function lacks formal expression and variable definitions;
- Some variables (like “k” for false alarm control) are used without proper definition.
A dedicated notation table might help.
- The conclusion should not only summarize the findings but also mention:
- Potential extensions (e.g., attention mechanisms, real-time SAR systems);
- Future directions (e.g., detecting even smaller or maneuvering targets);
- Generalization plans across different sea states or moving platforms.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsPlease check the attachment.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper proposes a multi-domain feature fusion method for small target radar detection in complex ocean backgrounds. It extracts features from four domains, compresses them with a stacked autoencoder, fuses using an MLP, and introduces an adjustable loss function to control false alarms. Experiments show superior performance on the IPIX dataset.
Specific comments:
(1) There are many references, but there is a lack of organized summary. It is recommended to systematically classify the existing methods, compare their advantages and disadvantages in the introduction, and highlight the positioning of this study.
(2) Although the false alarm rate is controlled by adjusting the loss function, the theoretical derivation and mathematical explanation of this mechanism are not sufficient. It is recommended to explain its principle more systematically.
(3) There is a lack of systematic discussion on the number of layers of the SAE structure, the hidden layer dimension, and the setting of the weight hyperparameters (α, β) in the loss function of the intra-domain network. It is recommended to add relevant ablation experiments or parameter sensitivity analysis.
(4) Although the current comparative experiments are compared with traditional methods and some fusion models, there is still a lack of comparative analysis with end-to-end deep neural networks (such as CNN or Transformer structures).
(5) The summary of the contributions of this paper in the conclusion is not in-depth enough. It is recommended to reiterate the main results and clarify the future research direction to enhance the foresight of the paper.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsMy comments have been well replied, there are no more questions.