Transformer Architecture for Micromotion Target Detection Based on Multi-Scale Subaperture Coherent Integration
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors1. The motivation and necessity of using the Transformer should be explained in section 1.
2. The detailed network parameters of the proposed method should be given besides Figure 3, including channel dimension of each block, etc.
3. The simulated dataset can be introduced more in detail, with statistical analysis of this dataset presented.
4. The results of the non-parametric search approaches mentioned in section 1 need to be added for comparison.
5. The paper mainly focuses on the detection of cylindrical space targets. How is the generalization performance on space moving targets with other shapes?
Author Response
Please see the attachment."
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper presents MsSCIFormer for target detection and motion parameter estimation. Starting with the micro-motion characteristics of CSTs, the authors designed a multi-scale subaperture processing module and a dual-attention mechanism to address the ARDU of scattering centers and improve performance in low SNR environments. Experimental results using simulation data show that, compared to other methods, this approach achieves the best evaluation metrics in both classification and regression tasks. Furthermore, extensive experiments have demonstrated the effectiveness of each component and the robustness of the overall method. However, there are a few issues that require further clarification:
1. Section 2 provides a detailed modeling of CST, but the authors should cite relevant references to support this part.
2. The use of symbols needs improvement: (1) Does RM6=-r share the same meaning as r in Equation (10)? (2) d is used to represent both "the base diameter of the cylindrical target" and as an index indicating different subaperture sizes later in the text. (3) Should Vitrd in Equation (16) actually be Vited? (4) Nf first mentioned in Section 3.2, is not defined until Section 3.3, where it is described as the feature dimension.
3. Why is a new dimension introduced in Section 3.2? Are the high-dimensional feature representations obtained through 2D convolution on the Nr*Nd plane? Please provide a more detailed explanation.
4. In Section 3.3, does the multi-scale approach generate D partitioning results for the same input, or does it randomly select a subaperture partitioning size from the set D for a given input?
5. In Section 3.3, the processing of the query matrix Q differs between Intra-SA and Inter-SA. Why is it that in the former,Q is not obtained through a linear transformation of X?
6. There seems to be a discrepancy between Figure 3 and the description in the text. If Q, K and V are the results of the linear layer transformation, shouldn’t their symbols be marked on the opposite side of the Linear layer?
7. In Section 4.2, it is necessary to supplement the comparison results of deep learning-based methods under low SNR conditions.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper has been carefully revised, with more details presented. I have no other questions, and suggest accepting it.
