PASTE: Physics-Aware Scattering Topology Embedding Framework for SAR Object Detection

Jiacheng Chen; Yuxuan Xiong; Haipeng Wang

cs.CVMar 16, 2026

PASTE: Physics-Aware Scattering Topology Embedding Framework for SAR Object Detection

Authors: Jiacheng Chen, Yuxuan Xiong, Haipeng Wang

Organizations: Key Laboratory for Information Science of Electromagnetic Waves (MoE), Fudan University, Shanghai 200433, China

Abstract

Current deep learning-based object detection for Synthetic Aperture Radar (SAR) imagery mainly adopts optical image methods, treating targets as texture patches while ignoring inherent electromagnetic scattering mechanisms. Though scattering points have been studied to boost detection performance, most methods still rely on amplitude-based statistical models. Some approaches introduce frequency-domain information for scattering center extraction, but they suffer from high computation cost and poor compatibility with diverse datasets. Thus, effectively embedding scattering topological information into modern detection frameworks remains challenging. To solve these problems, this paper proposes the Physics-Aware Scattering Topology Embedding Framework (PASTE), a novel closed-loop architecture for comprehensive scattering prior integration. By building the full pipeline from topology generation, injection to joint supervision, PASTE elegantly integrates scattering physics into modern SAR detectors. Specifically, it designs a scattering keypoint generation and automatic annotation scheme based on the Attributed Scattering Center (ASC) model to produce scalable and physically consistent priors. A scattering topology injection module guides multi-scale feature learning, and a scattering prior supervision strategy constrains network optimization by aligning predictions with scattering center distributions. Experiments on real datasets show that PASTE is compatible with various detectors and brings relative mAP gains of 2.9% to 11.3% over baselines with acceptable computation overhead. Visualization of scattering maps verifies that PASTE successfully embeds scattering topological priors into feature space, clearly distinguishing target and background scattering regions, thus providing strong interpretability for results.

Figures & tables

Fig. 1: Optical (top) and SAR (bottom) images of the same terminal area in Changi Airport. The right side of the figure highlights the same aircraft target using a red region and yellow dashed lines, respectively. It can be observed that the aircraft target appears as a continuous region in the optical image, whereas it is characterized by discrete bright patches in the SAR image.

Fig. 2: The overview architecture of the proposed PASTE that consists of three components: Generation, Supervision, and Injection. During training, the Scattering Keypoints Automatic Annotation (SKAA) for ”Generation” automatically generates scattering keypoints and forms extended annotation files that are fully compatible with the standard detection training pipeline. Then Scattering Prior Supervision Strategy (SPSS) for “Supervision” aligns the generated keypoints at the image scale and supervises the scattering map prediction network within the Scattering Topology Injection Module (STIM), guiding it to learn the scattering topology prior and predict target scattering regions. During inference, STIM for ”Injection” implicitly encodes the learned scattering topology prior and can directly predict the scattering map from the input image without the involvement of SKAA and SPSS, enabling feature fusion with existing detectors.

Fig. 4: Diagram of spatial scattering map prediction network and fusion flow

Fig. 5: Visualization of scattering keypoints for representative classes in the FAIR-CSAR training set. The first row shows examples from major aircraft categories, while the second row illustrates keypoint patterns of representative ship categories. All automatically generated scattering keypoints are marked with red dots in the figure.

Baseline Detector	Stage	Anchor	Justification for Selection
Rotated RetinaNet	single	anchor-based	Classic one-stage anchor-based baseline; widely used in rotated object detection.
GWD	single	anchor-free	Represents anchor-free regression with probabilistic localization loss.
RoI Transformer	two	anchor-based	Converts horizontal RoIs to rotated ones via learnable transformation; tests PASTE on geometry-aware proposal refinement.
Rotated FCOS	single	anchor-free	Dense anchor-free detector; tests PASTE on center-ness assignment.
Oriented R-CNN (ResNet-50)	two	anchor-based	Strong two-stage CNN baseline; demonstrates PASTE compatibility with standard backbones.
Oriented R-CNN (Swin-T)	two	anchor-based	Swin-Transformer variant; evaluates PASTE on Transformer backbones.

TABLE I: Baseline Detectors Adopted for PASTE Evaluation

Baseline Detector	[email protected]	[email protected] (+PASTE)	Gain	Relative gain (%)
Rotated RetinaNet	22.9	24	$\mathbf{1.1\uparrow}$	4.8%
GWD	26	27	$\mathbf{1.0\uparrow}$	3.8%
RoI Transformer	38	42.3	$\mathbf{4.3\uparrow}$	11.3%
Rotated FCOS	34.5	35.7	$\mathbf{1.2\uparrow}$	3.5%
Oriented R-CNN (ResNet-50)	39.2	43.3	$\mathbf{4.1\uparrow}$	10.5%
Oriented R-CNN (Swin-T)	41.3	44.4	$\mathbf{3.1\uparrow}$	7.5%

TABLE II: Performance Boost by PASTE on FAIR-CSAR (SL Mode)

Baseline Detector	[email protected]	[email protected] (+PASTE)	Gain	Relative gain (%)
Rotated RetinaNet	37.2	38.9	$\mathbf{1.7\uparrow}$	4.6%
GWD	39.4	40.6	$\mathbf{1.2\uparrow}$	3.0%
RoI Transformer	45.4	48.7	$\mathbf{3.3\uparrow}$	7.3%
Rotated FCOS	45.4	46.7	$\mathbf{1.3\uparrow}$	2.9%
Oriented R-CNN (ResNet-50)	44.8	48.3	$\mathbf{3.5\uparrow}$	7.8%
Oriented R-CNN (Swin-T)	50	52	$\mathbf{2.0\uparrow}$	4.0%

TABLE III: Performance Boost by PASTE on FAIR-CSAR (FSI Mode)

Fig. 6: Proposals (First stage output) visualization under three scenes. Row 1: inshore ships; Row 2: offshore ships; Row 3: terminal aircraft. Column (a): Ground-truth boxes; Column (b): Top-scoring proposals ( $\geq$ 0.5) generated by Oriented R-CNN equipped with PASTE; Column (c): Top-scoring proposals ( $\geq$ 0.5) generated by Oriented R-CNN without PASTE. All proposals are colored from cool to warm according to confidence score.

Fig. 7: Hit Rate of predicted Proposals under variant IoU threshold under three typical detection scenes: (a) Offshore ships; (b) Inshore Ships; (c) Terminal aircrafts.

Scenario	Proposal Precision (no PASTE)	Proposal Precision (with PASTE)
Offshore Ships	81.4%	93.1%
Inshore Ships	56.7%	59.5%
Terminal Aircrafts	63.8%	66.6%

TABLE IV: Precision of proposals in different scenarios

Fig. 8: Visualization of predicted spatial scattering map in different scenarios. The left column shows the test samples with ground-truth bounding boxes, while the right column presents the corresponding spatial scattering map predicted by the proposed model. From top to bottom: (a) aircraft detection at an airport, (b) oil tank detection in a ground industrial area, and (c) near-shore ship detection at a port. The highlighted regions in the scattering map correspond to the areas where the prediction network in STIM focuses on the key scattering structures.

Fig. 9: Predicted spatial scattering map on target region across representative categories.

Fig. 10: Visual examples of SKAA-Keypoints and AMP-Keypoints. Row 1: SKAA-Keypoints (Red); Row 2: AMP-Keypoints (Green); Column (a): Aircraft; Column (b): Ship; Column (c): Tank.

STIM	SPSS	Keypoints Type	[email protected]
✗	$\backslash$	$\backslash$	38
✓	✗	$\backslash$	39.9
✓	✓	AMP-Keypoints	40.6
✓	✓	SKAA-Keypoints	42.3

TABLE V: Ablation studies for different components of PASTE.

Number of Keypoints	[email protected]
3	41.6
6	41.4
9	43.3
12	42.6
15	42.1

TABLE VI: Ablation Study on Different Numbers of Keypoints

SPSS Weight $\lambda$		Gaussian Radius $\sigma$
$\lambda$	[email protected]	$\sigma$	[email protected]
0.1	42.5	1	43.3
0.5	42.2	2	42.4
1	43.3	4	42.1
2	41.8	8	41.1

TABLE VII: Ablation Studies on SPSS Hyperparameters

Stage	Metric	Baseline	PASTE	Overhead
Annotation Generation	Avg time per sample (ms)	–	413.5	–
Training	Model size (MB)	315.2	337.7	+7.1%
Training	Training time per epoch (s)	588.4	661.5	+12%
Inference	FLOPs (G)	211.44	274.3	+29.7%
Inference	Latency per image (ms)	61.7	72.4	+17.3%

TABLE VIII: Computational cost and memory overhead of proposed PASTE.

Object Detection

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

See all 624 papers

Jun 5, 2026

Physics-Driven Semantic Scattering Structure Understanding of Aircraft Target in SAR Images

Synthetic aperture radar (SAR) has become indispensable for target interpretation owing to its all-day and all-weather observation capability. In SAR target interpretation, electromagnetic scattering information provides a physically grounded cue beyond visual texture and has been widely exploited for target interpretation. However, existing methods remain dominated by local scattering center representations. Such unordered and component-agnostic representations are highly unstable for aircraft targets. As a result, physically existing components with weak scattering responses are often missed, resulting in the incomplete reconstructed topology structure. To address this limitation, we establish Semantic Scattering Structure Understanding as a new paradigm for SAR aircraft interpretation. Semantic scattering keypoints are defined to associate local electromagnetic responses with physically meaningful aircraft components, while visibility-aware attributes are introduced to retain weakly observable yet physically existed components. The keypoints are further organized into a stable semantic scattering structure. Build upon this, we propose S3U-SAR, a physics-driven framework to localize semantic scattering keypoints and construct the complete representation constrained by multi-dimensional physical priors containing scattering heterogeneity, rigid-body topology, speckle uncertainty. A confidence-gated joint supervision strategy is further introduced to alleviate optimization conflicts. We construct KP-SAR-Aircraft-1.0, the first fine-grained benchmark for semantic scattering structure understanding. Extensive experiments demonstrate that S3U-SAR achieves the best performance compared with baselines. Cross-category and cross-dataset evaluations further verify its robustness and transferability.

Yifei Yin, Xiaogang Yu, Hao Shi +2

School of Information and Electronics, Beijing Institute of Technology, Beijing 100081, China, and with the National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing · Beijing Institute of Remote Sensing Information, Beijing, China

Remote Sensing Image Understanding Physics-Informed ML Synthetic Aperture Radar

Ask PDF

Aug 19, 2026

SED-FOD: Scattering-Aware Expert Decomposition for Few-Shot Cross-Sensor SAR Object Detection

Synthetic aperture radar (SAR) object detection is an important part of remote sensing interpretation. However, because of variations in frequency band, resolution, background clutter, and target scattering responses, the performance of existing detectors often degrades when training and testing data are acquired from different SAR domains. Although domain adaptation methods offer a promising paradigm for solving this problem, most of them mainly pursue domain-invariant feature alignment and suppress sensor-dependent scattering characteristics that are useful for object detection. This problem becomes more challenging in few-shot scenarios, where only a few fully annotated target-domain SAR images are available. To address this issue, we propose a scattering-aware shared-specific feature decomposition framework for few-shot SAR domain adaptation object detection. We decompose detection features into a shared path and several soft-gated scattering-specific expert paths. The shared path learns transferable object structural information and is used for asymmetric domain alignment, while the scattering-specific experts adaptively compensate heterogeneous SAR responses. In addition, routing-domain auxiliary loss is introduced to encourage specific experts to capture sensor-dependent routing preferences, and an expert balancing loss is used to prevent routing collapse. Extensive experiments on four bidirectional heterogeneous SAR detection tasks between FARAD-X/FARAD-Ka and MiniSAR under different few-shot settings have been conducted and experimental results demonstrate that the proposed method achieves superior performance in both forward and reverse adaptation directions.

Shu Yang, Zhen Chen, Zhiyu Jiang +2

National Key Laboratory of Microwave Imaging Technology, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China, and also with the School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, China.

Synthetic Aperture Radar Few-Shot Domain Adaptation Cross-Domain Transfer Learning Mixture of Experts

Ask PDF

May 27, 2026

Lightweight SAR Ship Detection via Contrastive Distillation

Deep convolutional and transformer-based detectors achieve strong performance for SAR ship detection but are often computationally prohibitive for real-time or onboard deployment. Lightweight models offer improved efficiency yet struggle to capture the complex structural relationships inherent in SAR backscatter. Most existing SAR knowledge-distillation approaches rely on feature or logit matching, which enforces localized activation similarity while neglecting the geometric relationships among object representations. We propose a Structured Unified Relational knowledGE distillation framework for SAR Ship detection (SURGE) that transfers relational geometry from a powerful teacher detector to a compact student detector using a contrastive InfoNCE objective in a shared projection embedding space. To the best of our knowledge, this work presents the first transformer-based SAR ship detector knowledge distillation framework in SAR domain. The framework is architecture-agnostic in the sense that it provides a common region-level distillation interface for two-stage, one-stage and transformer-based detectors without modifying their deployed architectures. Experiments on the SSDD and HRSID benchmarks demonstrate that the proposed method yields substantial improvements for two-stage detectors, achieving up to 6.2 mAP and 8.0 AP75 gains over baseline student and even surpassing teacher performance

Surendar Devasundaram, Banafsheh Saber Latibari, Abhijit Mahalanobis

University of Arizona · Department of Electrical and Computer Engineering · Tucson, Arizona, USA

Efficient Object Detection Relational Knowledge Distillation Object Detection Synthetic Aperture Radar Knowledge Distillation Ship Detection

Ask PDF

Abstract

Object Detection

Explore similar work

Physics-Driven Semantic Scattering Structure Understanding of Aircraft Target in SAR Images

SED-FOD: Scattering-Aware Expert Decomposition for Few-Shot Cross-Sensor SAR Object Detection

Lightweight SAR Ship Detection via Contrastive Distillation