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

Fenghao Song; Shaojing Yang; Xi Zhou

cs.CVApr 5, 2026

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

Authors: Fenghao Song, Shaojing Yang, Xi Zhou

Organizations: Yunnan Normal University, No. 768, Jucheng Avenue, Chenggong District, Kunming 650500, China.

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.

Figures & tables

Fig. 1: SARES-DEIM overview. The backbone produces the multi-scale features, SARESMoE performs image-level Top- $k$ expert selection with an always-active Shared Expert, and SDEP injects P2 detail into the P3 pathway before the encoder and DETR-style decoder/head.

Model	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$P$ (%)	$R$ (%)	Params (M)	FLOPs (G)	FPS
Faster R-CNN [ 40 ]	–	88.0	82.1	–	136.7	147.8	–
SSD [ 41 ]	–	90.1	81.0	–	32.0	70.0	–
FCOS [ 42 ]	–	81.1	83.6	–	36.3	128.0	–
CSCF-Net [ 43 ]	–	91.5	90.7	–	2.61	9.4	–
SAR-D-FINE [ 44 ]	72.0	90.8	91.2	–	11.8	31.0	109.0
OptiSAR-Net [ 45 ]	65.4	90.2	–	–	2.7	11.8	90.6

TABLE I: Detection performance and efficiency comparison on HRSID [ 39 ] . Bold indicates the best detection metric among listed entries. Baseline entries marked with ∗ are locally reproduced under our implementation setting; unmarked entries are reported from the corresponding cited papers when available. SARES-DEIM is the proposed method evaluated under the same setting. Base denotes the direct DEIM-S [ 16 ] architectural baseline.

Model	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$P$ (%)	$R$ (%)	Params (M)	FLOPs (G)	FPS
Faster R-CNN [ 40 ]	–	86.3	80.7	–	136.7	147.8	–
SSD [ 41 ]	–	88.9	87.1	–	32.0	70.0	–
FCOS [ 42 ]	–	87.5	89.1	–	36.3	128.0	–
CoAM-Darknet [ 47 ]	–	96.0	93.7	95.3	65.8	123.5	–
YOLOv8-S ∗ [ 48 ]	66.1	97.1	94.5	93.6	9.8	23.6	212.8
YOLOv8-M ∗ [ 48 ]	66.2	95.5	91.5	90.8	23.2	67.8	204.8

TABLE II: Detection performance and efficiency comparison on SAR-Ship-Dataset [ 50 ] . Bold indicates the best detection metric among listed entries. Baseline entries marked with ∗ are locally reproduced under our implementation setting; unmarked entries are reported from the corresponding cited papers when available. SARES-DEIM is the proposed method evaluated under the same setting. Base denotes the direct DEIM-S [ 16 ] architectural baseline.

Model	Params (M)	FLOPs (G)	FPS (batch size 8)	Latency (ms, batch size 1)
YOLOv8-S [ 48 ]	9.8	23.6	212.8	11.2
YOLOv8-M [ 48 ]	23.2	67.8	204.8	11.5
YOLOv11-S [ 49 ]	9.4	21.5	243.9	10.1
YOLOv11-M [ 49 ]	20.0	68.2	172.4	13.0
RT-DETR-L [ 14 ]	28.8	104.8	106.4	66.2
D-FINE-S [ 17 ]	10.2	24.8	227.4	29.1

TABLE III: Efficiency comparison with batch-size-1 latency for locally reproduced models. FPS is measured with batch size 8, while latency is computed from batch-size-1 throughput under the same

640\times 640

input resolution on a single NVIDIA A40 GPU. Bold indicates the best efficiency value in each column; lower is better for Params, FLOPs, and latency, whereas higher is better for FPS.

Target Test Set	Method	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$P$ (%)	$R$ (%)	TP	FP	FN
SAR-Ship-Dataset [ 50 ]	DEIM-S (Base) [ 16 ]	15.9	49.1	28.2	62.9	9,628	24,530	5,670
SAR-Ship-Dataset [ 50 ]	SARES-DEIM	17.2	54.7	34.6	67.0	10,248	19,352	5,050
SSDD [ 51 ]	DEIM-S (Base) [ 16 ]	34.3	70.3	47.4	75.2	401	445	132
SSDD [ 51 ]	SARES-DEIM	35.2	75.1	51.1	80.1	427	408	106

TABLE IV: Source-only cross-dataset evaluation of DEIM-S [ 16 ] and SARES-DEIM with explicit prediction-to-target category mapping from HRSID [ 39 ] to SAR-Ship-Dataset [ 50 ] and SSDD [ 51 ] . AP uses the official COCO evaluation [ 52 ] .

P

R

, and the error counts use a confidence threshold of 0.25 and an IoU threshold of 0.50. AP,

P

, and

R

are reported in percent. Bold indicates the better result for each target dataset; lower is better for FP and FN.

Target Test Set	Confidence	DEIM-S $P$	DEIM-S $R$	SARES-DEIM $P$	SARES-DEIM $R$
SAR-Ship-Dataset [ 50 ]	0.10	7.4	69.7	13.0	72.6
	0.25	28.2	62.9	34.6	67.0
	0.50	55.7	55.1	60.1	59.6
SSDD [ 51 ]	0.10	11.2	82.4	20.1	86.5
	0.25	47.4	75.2	51.1	80.1
	0.50	80.4	63.2	82.1	70.4

TABLE V: Shared-confidence-threshold analysis for DEIM-S [ 16 ] and SARES-DEIM on SAR-Ship-Dataset [ 50 ] and SSDD [ 51 ] at an IoU threshold of 0.50. All values are reported in percent; bold indicates the better result at each shared threshold.

Fig. 2: Confidence-ranked precision–recall curves after explicit category mapping: (a) HRSID [ 39 ]

\rightarrow

SAR-Ship-Dataset [ 50 ] and (b) HRSID

\rightarrow

SSDD [ 51 ] .

Fig. 3: TP/FP/FN comparison at a confidence threshold of 0.25 and an IoU threshold of 0.50: (a) HRSID [ 39 ]

\rightarrow

SAR-Ship-Dataset [ 50 ] and (b) HRSID

\rightarrow

SSDD [ 51 ] .

Scenario	Method	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$P$ (%)	$R$ (%)
Inshore	DEIM-S [ 16 ] (Base)	74.7	92.3	90.3	86.3
Inshore	SARES-DEIM	75.8	93.2	90.9	87.7
Offshore	DEIM-S [ 16 ] (Base)	90.3	99.4	98.8	98.6
Offshore	SARES-DEIM	91.0	99.4	99.1	98.6

TABLE VI: Scene-wise performance comparison on the official HRSID [ 39 ] inshore/offshore subsets. All detection metrics are reported in percent, and bold indicates the better result within each scene group; ties are both bold. Base denotes the direct DEIM-S [ 16 ] architectural baseline.

SDEP	SARESMoE Stages			$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$AP_{S}$ (%)	$AP_{M}$ (%)	$AP_{L}$ (%)	$P$ (%)	$R$ (%)
SDEP	P3	P4	P5	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$AP_{S}$ (%)	$AP_{M}$ (%)	$AP_{L}$ (%)	$P$ (%)	$R$ (%)
$\times$	$\times$	$\times$	$\times$	73.0	92.0	73.5	75.0	32.5	92.4	84.4
$\times$	$\checkmark$	$\times$	$\times$	74.7	92.9	76.0	76.2	34.1	92.5	87.4
$\times$	$\checkmark$	$\checkmark$	$\times$	72.9	92.8	73.6	72.7	31.0	91.4	85.9
$\times$	$\checkmark$	$\checkmark$	$\checkmark$	75.3	93.6	76.0	76.2	34.3	92.0	87.1
$\checkmark$	$\times$	$\times$	$\times$	74.9	92.9	76.2	76.3	33.1	93.1	87.3

TABLE VII: Module-level ablation on the HRSID dataset [ 39 ] with scale-specific AP metrics. The SARESMoE module is incrementally applied to the spatial pyramid stages (P3, P4, and P5). Bold indicates the best value in each detection-metric column.

Routing Strategy	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$AP_{S}$ (%)	$AP_{M}$ (%)	$AP_{L}$ (%)	$P$ (%)	$R$ (%)
Standard MLP Router	75.2	93.4	75.8	76.2	24.6	86.6	89.6
Frequency Branch Only	75.3	93.6	76.2	75.2	20.6	87.7	89.9
Spatial Branch Only	75.7	93.8	76.4	75.8	28.7	87.9	90.2
Dual-Branch Fusion (Ours)	76.4	93.8	77.0	76.8	34.6	93.1	88.0

TABLE VIII: Ablation study of distinct learnable router architectures within SARESMoE on HRSID [ 39 ] with scale-specific AP metrics. Bold indicates the best performance in each detection-metric column.

Variant	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$P$ (%)	$R$ (%)	Params (M)	FLOPs (G)
Uniform All-Expert Aggregation	75.6	93.4	93.1	87.9	87.6	128.8
Fixed Top-2 Routing	75.4	93.2	92.1	88.4	90.7	75.3
Dynamic Top-2 Routing (Ours)	76.4	93.8	93.1	88.0	90.8	72.6

TABLE IX: Comparison of routing policies under the 4/8/16 heterogeneous expert-bank configuration on HRSID [ 39 ] . The routing policies are defined in the accompanying text; the table reports only quantitative results. Bold indicates the best value in each reported column; lower is better for Params and FLOPs.

Routing Policy	Seed	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$P$ (%)	$R$ (%)
Fixed Top-2 Routing	0	75.4	93.2	92.1	88.4
	13	75.1	92.8	93.2	87.3
	21	75.1	93.4	92.5	88.4
	42	75.4	93.0	93.1	87.7
	Mean $\pm$ SD	$75.25\pm 0.17$	$93.10\pm 0.26$	$92.73\pm 0.52$	$87.95\pm 0.54$
Dynamic Top-2 Routing	0	76.4	93.8	93.1	88.0

TABLE X: Four-seed comparison of Fixed Top-2 Routing and Dynamic Top-2 Routing on HRSID [ 39 ] . Mean

\pm

sample standard deviation is reported in percentage points; bold indicates the better mean in each detection-metric column.

Fusion Strategy (P2 $\rightarrow$ P3)	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$AP_{S}$ (%)	$AP_{M}$ (%)	$AP_{L}$ (%)	$P$ (%)	$R$ (%)
Baseline (No P2 Fusion)	75.3	93.6	76.0	76.2	34.3	92.0	87.1
Standard Strided Downsampling	74.4	93.2	75.3	74.4	24.3	86.3	89.6
Strip Convolution [ 53 ]	74.1	93.2	75.0	74.7	28.1	86.8	89.4
SDEP (Space-to-Depth)	76.4	93.8	77.0	76.8	34.6	93.1	88.0

TABLE XI: Comparison of different P2-to-P3 fusion strategies on HRSID [ 39 ] with scale-specific AP metrics. All configurations are evaluated on top of the SARESMoE baseline. Bold indicates the best performance in each detection-metric column.

Configuration	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$AP_{S}$ (%)	$AP_{M}$ (%)	$AP_{L}$ (%)	$P$ (%)	$R$ (%)
Shared Expert Only	74.5	92.7	75.8	76.7	32.2	92.1	86.6
P3: Spatial Expert Only	73.8	92.4	75.2	75.3	28.6	92.8	85.6
P3: Wavelet Expert Only	74.1	92.7	75.4	75.8	33.3	92.5	87.3
P4: Frequency Expert Only	73.9	92.3	75.1	75.2	36.9	91.9	86.9
P4: Hybrid Expert Only	74.2	92.8	75.6	74.9	29.6	92.3	86.6
P5: Frequency Expert Only	74.6	92.6	75.7	77.1	31.1	92.9	86.8

TABLE XII: Expert composition ablation on HRSID [ 39 ] with scale-specific AP metrics. All models are evaluated without SDEP to isolate the MoE mechanism. The Shared Expert is uniformly active across all runs. For single-expert ablation rows, the specified level utilizes only the designated expert, while the unmentioned pyramid levels retain the default full SARESMoE expert configuration. Bold indicates the best performance in each detection-metric column.

Expert Number	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$AP_{S}$ (%)	$AP_{M}$ (%)	$AP_{L}$ (%)	Params (M)	FLOPs (G)
2/2/2	74.7	93.1	75.9	75.9	33.5	37.6	72.6
4/4/4	74.8	92.9	75.9	76.5	32.1	46.4	72.6
8/8/8	75.1	93.0	76.4	76.0	35.5	63.9	72.6
16/16/16	74.8	93.0	76.1	75.6	30.2	99.1	72.6
Ours (4/8/16)	76.4	93.8	77.0	76.8	34.6	90.8	72.6

TABLE XIII: Ablation study of SARESMoE expert number on HRSID [ 39 ] . All settings use Top-2 sparse activation; only the number of stored experts changes. Bold indicates the best detection metric in each column.

Gating Setting	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$AP_{S}$ (%)	$AP_{M}$ (%)	$AP_{L}$ (%)	Params (M)	FLOPs (G)
Top-1	75.2	93.6	76.5	75.9	37.7	90.8	67.1
Top-All	75.0	92.9	76.2	76.8	35.4	90.8	126.0
Top-2 (Ours)	76.4	93.8	77.0	76.8	34.6	90.8	72.6

TABLE XIV: Ablation study of Top-

k

gating sparsity in SARESMoE on HRSID [ 39 ] . The expert bank is kept unchanged, while only the number of activated sparse experts is varied. Bold indicates the best detection metric in each column; lower FLOPs is preferred.

SDEP Insertion Level	$mAP_{50:95}$ (%)	$mAP_{50}$ (%)	$AP_{S}$ (%)	$AP_{M}$ (%)	$AP_{L}$ (%)	Params (M)	FLOPs (G)
No SDEP	75.3	93.6	76.0	76.2	34.3	90.2	63.4
P2 $\rightarrow$ P4	74.1	92.8	75.3	75.9	41.8	92.9	68.4
P2 $\rightarrow$ P5	74.0	92.4	75.3	75.4	28.5	91.2	67.5
P2 $\rightarrow$ P3 (Ours)	76.4	93.8	77.0	76.8	34.6	90.8	72.6

TABLE XV: Ablation study of SDEP insertion levels on HRSID [ 39 ] . P2 features are injected into different pyramid stages and evaluated with scale-specific AP metrics. Bold indicates the best value in each column; lower is better for Params and FLOPs.

Fig. 4: Qualitative bounding-box comparison on representative HRSID [ 39 ] samples. The figure compares representative YOLO-series, DETR-series, and SARES-DEIM detection outputs under challenging SAR ship-detection scenarios.

Fig. 5: Qualitative bounding-box comparison on representative SAR-Ship-Dataset [ 50 ] samples. The figure compares representative YOLO-series, DETR-series, and SARES-DEIM detection outputs under different SAR ship appearances and backgrounds.

Fig. 6: Qualitative comparison between the ground truth (GT), the DEIM-S baseline (Base), and the proposed SARES-DEIM (Ours) on representative HRSID [ 39 ] samples. Yellow ellipses indicate missed detections, while red ellipses indicate false alarms. The focused comparison shows that SARES-DEIM can reduce several missed detections on small or weak targets and suppress some clutter-induced false positives compared with the baseline.

Fig. 7: Expert-type activation statistics of SARESMoE on inshore and offshore HRSID [ 39 ] subsets. The percentages indicate normalized frequencies of Top- $k$ sparse expert-selection events at P3, P4, and P5.

Fig. 8: Representative paired routing cases on official HRSID [ 39 ] inshore and offshore subsets. Each pair contains one inshore and one offshore SAR image, and the badges list the Top- $k$ selected sparse experts and their normalized routing weights at P3, P4, and P5. The visualization reflects image-level and pyramid-level expert selection rather than dense pixel-wise routing.

Fig. 9: Expert-level CAM visualizations on HRSID [ 39 ] . All configurations exclude SDEP: (a) all sparse experts replaced with the Shared Expert; (b) P3 Spatial only; (c) P3 Wavelet only; (d) P4 Frequency only; (e) P4 Hybrid only; (f) P5 Frequency only; (g) P5 Hybrid only; (h) Uniform All-Expert Aggregation; and (i) full SARESMoE. Unmentioned pyramid levels retain the default full SARESMoE expert configuration. Warmer colors indicate stronger feature responses.

Fig. 10: Module-level ablation visualizations on HRSID [ 39 ] corresponding to Table VII : (a) DEIM-S baseline; (b) DEIM-S + SARESMoE; (c) DEIM-S + SDEP; and (d) full SARES-DEIM. Each column displays the detection output (top) and corresponding CAM heatmap (bottom). Warmer colors indicate stronger feature responses.

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

Prompting Foundation Models for Zero-Shot Ship Instance Segmentation in SAR Imagery
Apr 20, 2026Islam Mansour, Francescopaolo Sica, Michael SchmittInstance Segmentation Promptable Image Segmentation
Department of Aerospace Engineering, University of the Bundeswehr Munich, Neubiberg, Germany
Lightweight SAR Ship Detection via Contrastive Distillation
May 27, 2026Surendar Devasundaram, Banafsheh Saber Latibari, Abhijit MahalanobisEfficient Object Detection Relational Knowledge Distillation
University of Arizona · Department of Electrical and Computer Engineering · Tucson, Arizona, USA
RSTNet: Enhancing Small-Target Recognition in Noisy SAR Imagery via Robust Feature Learning and Distribution-Aware Regression
Feb 27, 2026Xiaojing Zhao, Shiyang Li, Zenan Chu +5Synthetic Aperture Radar Small Object Detection
College of Mathematics and Information Science, Anyang Institute of Technology, Anyang, 455000, China · School of Computer Science and Information Engineering, Anyang Institute of Technology, Anyang, 455000, China · College of Electronic Information and Optical Engineering, Nankai University, Tianjin, 300350, China +3

Apr 20, 2026J/K move · Enter open · S save

Prompting Foundation Models for Zero-Shot Ship Instance Segmentation in SAR Imagery

Islam Mansour, Francescopaolo Sica, Michael Schmitt

Department of Aerospace Engineering, University of the Bundeswehr Munich, Neubiberg, Germany

Instance Segmentation Promptable Image Segmentation Remote Sensing Synthetic Aperture Radar

Open paper AskPDF arXiv

Synthetic Aperture Radar (SAR) plays a critical role in maritime surveillance, yet deep learning for SAR analysis is limited by the lack of pixel-level annotations. This paper explores how general-purpose vision foundation models can enable zero-shot ship instance segmentation in SAR imagery, eliminating the need for pixel-level supervision. A YOLOv11-based detector trained on open SAR datasets localizes ships via bounding boxes, which then prompt the Segment Anything Model 2 (SAM2) to produce instance masks without any mask annotations. Unlike prior SAM-based SAR approaches that rely on fine tuning or adapters, our method demonstrates that spatial constraints from a SAR-trained detector alone can effectively regularize foundation model predictions. This design partially mitigates the optical-SAR domain gap and enables downstream applications such as vessel classification, size estimation, and wake analysis. Experiments on the SSDD benchmark achieve a mean IoU of 0.637 (89% of a fully supervised baseline) with an overall ship detection rate of 89.2%, confirming a scalable, annotation-efficient pathway toward foundation-model-driven SAR image understanding.

Abstract

Object Detection

Explore similar work

Prompting Foundation Models for Zero-Shot Ship Instance Segmentation in SAR Imagery

Lightweight SAR Ship Detection via Contrastive Distillation

RSTNet: Enhancing Small-Target Recognition in Noisy SAR Imagery via Robust Feature Learning and Distribution-Aware Regression