SARES-DEIM: Sparse Mixture-of-Experts Meets DETR for Clutter-Aware SAR Ship Detection

arXiv:2604.04127 · cs.CV · Submitted 2026-04-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "SARES-DEIM: Sparse Mixture-of-Experts Meets DETR for Clutter-Aware SAR Ship Detection".

Jane: Here's the gist of the paper. The gist: SARES-DEIM is a domain-aware detection framework grounded in the DEtection TRansformer (DETR) paradigm that addresses challenges in SAR ship detection by incorporating a SAR-aware Expert Selection Mixture-of-Experts module and a Space-to-Depth Enhancement Pyramid neck to achieve superior performance on challenging maritime scenarios.

Tom: How it works SARES-DEIM is designed as an end-to-end set prediction paradigm, where the backbone extracts multi-scale feature maps, and then the model adapts to SAR imagery through domain integration The framework integrates several key components to handle the unique backscattering characteristics of SAR imagery, such as coherent speckle noise and coastal clutter The core of the adaptation lies in the SARESMoE module, which employs a sparse gating mechanism to selectively route features toward specialized frequency and wavelet experts This sparsely-activated architecture effectively filters speckle noise and semantic clutter while maintaining high computational efficiency The SAR-aware routing mechanism generates a global context descriptor via global average pooling and projects it to a probability simplex using the formulation z = Wr · Pavg(X) + br, π = Softmax(z/τ) The expert design is physically explainable and orthogonal, assigning experts based on pyramid level requirements Specifically, Wavelet & Spatial Experts are assigned to P3 for diminutive ships highly sensitive to speckle, employing WTConv

31: for learnable soft-thresholding and GhostNet modules

32: for structural details Frequency & Hybrid Experts are assigned to P4/P5 to handle complex sea clutter, utilizing mechanisms from FADC

33: to construct Frequency Experts and combining parallel branches for Hybrid Experts How it works The Space-to-Depth Enhancement Pyramid (SDEP) neck is introduced to preserve high-resolution spatial cues from shallow stages, significantly improving the localization of small targets This mechanism addresses the deep semantic loss inherent in DETR architectures by explicitly harvesting high-resolution information from the P2 layer (1/4 scale) The SDEP neck utilizes Space-to-Depth Convolution (SPDConv)

21: to perform lossless downsampling, reshaping the feature map F2 as SPD(F2) = Reshape F2; H2 2, W2 2, 4C2 This operation halves the spatial resolution to match the P3 stride while strictly preserving all spatial information in the expanded channel dimension After this transformation, a 1 × 1 convolution projects the expanded channels back to dimension d and fuses it with P3 features: P3 = Conv1×1(SPD(F2)) + F3 This enriched P3 then participates in the standard bidirectional fusion pathway to produce the final pyramid How it works The final detection set Yˆ is produced by a DETR-style decoder

9: processing N object queries against these pyramid features, yielding prediction logits and normalized bounding boxes The synergy between SDEP and SARESMoE is crucial for performance How it works The effectiveness of the proposed modules is validated through extensive ablation studies on the HRSID dataset The full SARESDEIM model achieves the optimal mAP50:95 of 76.4% and the highest APsmall of 77.0% This superior performance is confirmed by the synergistic relationship where SDEP provides essential structural details for small-target localization, while SARESMoE ensures these precise cues are shielded from SAR-specific background interference How it works The SAR-Aware Routing Mechanism analysis shows that the Dual-Branch Fusion tightly integrates both frequency and spatial domain analyses, recovering Precision to a peak of 93.1% and achieving the highest overall mAP50:95 (76.4%) This confirms that simultaneous extraction of spectral energy distributions and spatial contrast is essential for accurate expert assignment in heterogeneous maritime scenes How it works Qualitative detection evaluation demonstrates that the full SARES-DEIM model produces bounding boxes that exhibit a much tighter fit to the target hulls compared to the Baseline, visually confirming that its architectural enhancements facilitate more accurate coordinate regression Furthermore, in challenging scenarios, SARES-DEIM demonstrates markedly superior robustness compared to the Baseline by effectively suppressing clutter-induced errors How it works The module-level ablation visualizations show that integrating SDEP into the baseline without any MoE modules yields a notable improvement of 1.9% in mAP50:95 (74.9% vs.

Jane: 73.0%) and 2.9% in Recall (87.3%) The superiority of Space-to-Depth (SDEP) is confirmed because it entirely bypasses the aliasing issues inherent in strided operations, pushing precision to a peak of 93.1% while achieving the highest mAP50:95 (76.4%) How it works The expert composition analysis indicates that the full SARESMoE configuration achieves the optimal trade-off by leveraging complementary strengths, as it surpasses both SARESMoE alone and SDEP alone The visual CAM heatmaps for the full SARES-DEIM model exhibit maximum target-centric focus with nearly zero activation in background areas, confirming a strong complementary relationship How it works The paper concludes that SARES-DEIM establishes a new state-of-the-art on HRSID across all evaluation metrics, achieving mAP50:95 of 76.4% and mAP50 of 93.8% This performance surpasses existing YOLO-series, DETR variants, and SAR-specific detectors across all evaluation metrics The domain-aware routing mechanism provides stronger representational capacity than task-specific architectural modifications alone The balanced and high scores in both Precision (96.7%) and Recall (96.

Paper summary: Jane: So, wrapping up on SARES-DEIM: the authors are really pushing this framework as a robust DETR-based approach that bridges general object detection with specific SAR domain requirements <ref:2604.04127#pg2>. What's the bigger picture here for how we think about building AI models for sensors?

Tom: It’s about showing that you can get strong results by adding domain-specific enhancements, like SARESMoE and SDEP, on top of a general architecture like DETR, instead of just trying to force a standard model to understand SAR data.

Lu: I think the key contribution is demonstrating that the domain-aware routing mechanism provides better representational capacity than just tweaking the base architecture with task-specific modifications 3pg <ref:2604.04127#pg3>. It’s about how you structure the model to look at both frequency and spatial domains at once.

Meng: From a production standpoint, it proves that even in a resource-constrained environment, you can get high accuracy if you design your feature selection process to be selective rather than brute force <ref:2604.04127#pg3>.

Jane: So, if we’re talking about what this means for the world of AI research right now, is it that we should focus more on these multi-expert systems instead of just bigger models?

Tom: It’s less about size and more about intelligent routing. The work confirms that carefully designed components that filter out noise and preserve fine details are what deliver the real gains in difficult sensing tasks <ref:2604.04127#pg1>.

Conclusion: Tom: So, SARES-DEIM is this framework that uses experts to filter out clutter in SAR imagery, right?

Jane: Yeah, so we talked about how it uses a mixture of experts and a pyramid structure to handle those tricky speckle noises and coastal stuff.

Lu: The core idea is that you use these sparse routing mechanisms to selectively send features toward the right specialist—wavelet for small ships, frequency experts for sea clutter.

Meng: For deployment, I’m thinking about the computational cost of all that expert switching; we need to know if it actually runs fast on real hardware.

Lalam: The authors say this setup achieves mAP50 of seventy-six point four percent and mAP50 of ninety-three point eight percent on the HRSID dataset.

Tom: Those are solid numbers, but what do those numbers mean for someone just listening to the show?

Jane: It means that compared to other methods, this framework is doing a better job finding small ships in noisy SAR pictures.

Lu: Specifically, they say the synergy between the pyramid structure and the expert selection is what really boosts precision to ninety-three point one percent.

Meng: That’s good, but does it actually handle a huge number of targets without slowing down? What about scaling up?

Lalam: The paper shows that even with a batch size of eight on an NVIDIA A40 GPU, the active computation is around seventy-two gflops.

Tom: Seventy-two gflops is manageable for inference, which is good for production systems. But what about the hard part—the training phase?

Jane: The authors focused heavily on showing that this domain-aware routing gives better representational capacity than just changing the base architecture with task-specific modifications.

Lu: They claim it surpasses both using SARESMoE alone and using SDEP alone, which suggests that combining these two components is what actually works best for the SAR data.

Meng: So, if I were trying to deploy this right now, would I focus on the expert selection part or the depth enhancement neck first?

Jane: It seems like both are critical; SDEP handles the spatial details for small targets, and SARESMoE shields those details from interference.

Lalam: The authors conclude that SARES-DEIM sets a new state-of-the-art on HRSID across all evaluation metrics.

Tom: That’s the main conclusion, a new top score on a challenging dataset. It shows that domain awareness really matters when you're dealing with complex sensor data like SAR.

Jane: And it points toward future work in how we can design these expert systems to be even more efficient and adaptable across different types of maritime clutter.

Fenghao Song, Shaojing Yang, Xi Zhou

Yunnan Normal University

cs.CV

Submitted: 2026-04-05

Updated: 2026-10-20

Comments: 17 pages, 10 figures, 15 tables. Accepted for publication in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)

Code: https://github.com/ultralytics/ultralytics

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 91/100

The gist: The gist: SARES-DEIM is a domain-aware detection framework grounded in the DEtection TRansformer (DETR) paradigm that addresses challenges in SAR ship detection by incorporating a SAR-aware Expert

Key concepts

SARESMoE
This is the core adaptation module that uses a sparse gating mechanism to selectively send features to different specialized experts. These experts are tuned for specific aspects of SAR data, such as wavelet patterns or frequency information. This allows the model to efficiently filter out noise and clutter while focusing computational power only where it is most needed.
Space-to-Depth Enhancement Pyramid (SDEP) neck
The SDEP neck is a mechanism designed to preserve high-resolution spatial details from early stages of the network. It uses Space-to-Depth Convolution to downsample features in a way that keeps all spatial information intact, even when reducing resolution. This is crucial for accurately locating small targets by maintaining fine structural cues.
DETR paradigm
The framework is grounded in the Detection Transformer (DETR) architecture, which uses an end-to-end set prediction approach. Instead of traditional bounding box regression, DETR uses object queries to directly predict the final set of detected objects. This structure provides a strong foundation for learning complex spatial relationships within the SAR image.
Domain-aware detection
This refers to adapting a model specifically for the unique characteristics of Synthetic Aperture Radar (SAR) imagery, such as coherent speckle noise and coastal clutter. By incorporating domain-specific modules like SARESMoE, the framework moves beyond general detectors to understand how SAR data behaves, leading to better performance on maritime scenes.

Terminology

Summary

The gist: SARES-DEIM is a domain-aware detection framework grounded in the DEtection TRansformer (DETR) paradigm that addresses challenges in SAR ship detection by incorporating a SAR-aware Expert Selection Mixture-of-Experts module and a Space-to-Depth Enhancement Pyramid neck to achieve superior performance on challenging maritime scenarios.

How it works

SARES-DEIM is designed as an end-to-end set prediction paradigm, where the backbone extracts multi-scale feature maps, and then the model adapts to SAR imagery through domain integration The framework integrates several key components to handle the unique backscattering characteristics of SAR imagery, such as coherent speckle noise and coastal clutter

The core of the adaptation lies in the SARESMoE module, which employs a sparse gating mechanism to selectively route features toward specialized frequency and wavelet experts This sparsely-activated architecture effectively filters speckle noise and semantic clutter while maintaining high computational efficiency The SAR-aware routing mechanism generates a global context descriptor via global average pooling and projects it to a probability simplex using the formulation z = Wr · Pavg(X) + br, π = Softmax(z/τ)

The expert design is physically explainable and orthogonal, assigning experts based on pyramid level requirements Specifically, Wavelet & Spatial Experts are assigned to P3 for diminutive ships highly sensitive to speckle, employing WTConv [31] for learnable soft-thresholding and GhostNet modules [32] for structural details Frequency & Hybrid Experts are assigned to P4/P5 to handle complex sea clutter, utilizing mechanisms from FADC [33] to construct Frequency Experts and combining parallel branches for Hybrid Experts

How it works

The Space-to-Depth Enhancement Pyramid (SDEP) neck is introduced to preserve high-resolution spatial cues from shallow stages, significantly improving the localization of small targets This mechanism addresses the deep semantic loss inherent in DETR architectures by explicitly harvesting high-resolution information from the P2 layer (1/4 scale)

The SDEP neck utilizes Space-to-Depth Convolution (SPDConv) [21] to perform lossless downsampling, reshaping the feature map F2 as SPD(F2) = Reshape F2; H2 2, W2 2, 4C2 This operation halves the spatial resolution to match the P3 stride while strictly preserving all spatial information in the expanded channel dimension After this transformation, a 1 × 1 convolution projects the expanded channels back to dimension d and fuses it with P3 features: P3 = Conv1×1(SPD(F2)) + F3 This enriched P3 then participates in the standard bidirectional fusion pathway to produce the final pyramid

How it works

The final detection set Yˆ is produced by a DETR-style decoder [9] processing N object queries against these pyramid features, yielding prediction logits and normalized bounding boxes The synergy between SDEP and SARESMoE is crucial for performance

How it works

The effectiveness of the proposed modules is validated through extensive ablation studies on the HRSID dataset The full SARESDEIM model achieves the optimal mAP50:95 of 76.4% and the highest APsmall of 77.0% This superior performance is confirmed by the synergistic relationship where SDEP provides essential structural details for small-target localization, while SARESMoE ensures these precise cues are shielded from SAR-specific background interference

How it works

The SAR-Aware Routing Mechanism analysis shows that the Dual-Branch Fusion tightly integrates both frequency and spatial domain analyses, recovering Precision to a peak of 93.1% and achieving the highest overall mAP50:95 (76.4%) This confirms that simultaneous extraction of spectral energy distributions and spatial contrast is essential for accurate expert assignment in heterogeneous maritime scenes

How it works

Qualitative detection evaluation demonstrates that the full SARES-DEIM model produces bounding boxes that exhibit a much tighter fit to the target hulls compared to the Baseline, visually confirming that its architectural enhancements facilitate more accurate coordinate regression Furthermore, in challenging scenarios, SARES-DEIM demonstrates markedly superior robustness compared to the Baseline by effectively suppressing clutter-induced errors

How it works

The module-level ablation visualizations show that integrating SDEP into the baseline without any MoE modules yields a notable improvement of 1.9% in mAP50:95 (74.9% vs. 73.0%) and 2.9% in Recall (87.3%) The superiority of Space-to-Depth (SDEP) is confirmed because it entirely bypasses the aliasing issues inherent in strided operations, pushing precision to a peak of 93.1% while achieving the highest mAP50:95 (76.4%)

How it works

The expert composition analysis indicates that the full SARESMoE configuration achieves the optimal trade-off by leveraging complementary strengths, as it surpasses both SARESMoE alone and SDEP alone The visual CAM heatmaps for the full SARES-DEIM model exhibit maximum target-centric focus with nearly zero activation in background areas, confirming a strong complementary relationship

How it works

The paper concludes that SARES-DEIM establishes a new state-of-the-art on HRSID across all evaluation metrics, achieving mAP50:95 of 76.4% and mAP50 of 93.8% This performance surpasses existing YOLO-series, DETR variants, and SAR-specific detectors across all evaluation metrics The domain-aware routing mechanism provides stronger representational capacity than task-specific architectural modifications alone The balanced and high scores in both Precision (96.7%) and Recall (96.

Improvements for AI systems

  1. SAR-aware Expert Selection Mixture-of-Experts (SARESMoE) for Domain Adaptation: The system can dynamically route features to specialized experts based on local scattering context, effectively suppress[ing] speckle noise and semantic clutter through domain-specific feature refinement.

  2. Space-to-Depth Enhancement Pyramid (SDEP) for Small Target Localization: The system can explicitly harvest high-resolution information from the P2 layer (1/4 scale) using a lossless downsampling mechanism to preserve the structural integrity of small targets, leading to improved localization of diminutive ships.

  3. Dual-Branch Expert Selection for Robust Feature Discrimination: By combining frequency and spatial domain analyses, the system can achieve a peak of 93.1% Precision, demonstrating that simultaneous extraction of spectral energy distributions and spatial contrast is essential for accurate expert assignment in heterogeneous maritime scenes.

  4. Lossless Spatial Cue Transmission via Space-to-Depth Convolution: The system can utilize the SPD operation, which rearranges spatial pixels into the channel dimension without any information loss, to entirely bypass the aliasing issues inherent in strided operations when injecting shallow-stage features into the semantic pyramid.

Abstract

Ship detection in Synthetic Aperture Radar (SAR) imagery is challenged by coherent speckle noise, complex coastal clutter, and the prevalence of small-scale targets. Conventional detectors, primarily designed for optical imagery, often exhibit limited adaptability to SAR-specific degradation and suffer from the loss of fine-grained ship signatures during spatial downsampling. To address these limitations, we propose SARES-DEIM, a domain-aware detection framework grounded in the DEtection TRansformer (DETR) paradigm. Central to our approach is SARESMoE (SAR-aware Expert Selection Mixture-of-Experts), a module leveraging a sparse gating mechanism to selectively route features toward specialized wavelet, spatial, and frequency-oriented experts. This sparsely activated architecture enhances clutter-aware representation while keeping the active inference cost controlled. Furthermore, we introduce the Space-to-Depth Enhancement Pyramid (SDEP) neck to preserve high-resolution spatial cues from shallow stages and improve the localization of small targets. Experiments on HRSID and SAR-Ship-Dataset show that SARES-DEIM achieves competitive results across both public benchmarks, with more prominent advantages on the challenging HRSID dataset. Specifically, SARES-DEIM achieves 76.4% mAP50:95 and 93.8% mAP50 on HRSID, while keeping the active computation at 72.6 GFLOPs and 136.5 FPS on a single NVIDIA A40 GPU with batch size 8; the batch-size-1 latency is 37.6 ms.

Sources

Related papers