Search papers, labs, and topics across Lattice.
This paper introduces a diffusion model-based approach for multi-scale ship detection in SAR imagery, addressing limitations of existing CFAR and deep learning methods in complex environments. The model learns to recover ground-truth bounding boxes from noisy initial boxes through a Markov chain, enabling anchor-free detection. A segmentation-derived land-sea mask is incorporated to reduce false positives. Experiments on SAR ship datasets demonstrate competitive accuracy and robustness compared to existing methods.
Ditch the anchors: diffusion models can directly detect multi-scale ships in SAR images with competitive accuracy and robustness, even in complex coastal environments.
Synthetic Aperture Radar (SAR) enables all-weather and all-day imaging, playing a critical role in military and civilian applications. Ship detection, as a fundamental task in SAR image interpretation, supports downstream applications including attitude estimation and target tracking. Although existing methods address multi-scale ship detection through traditional strategies like Constant False Alarm Rate (CFAR) or deep learningbased feature fusion architectures, they often suffer from limited robustness in complex scenarios (e.g., ports and nearshore areas) and high computational complexity. Recently, diffusion models have shown remarkable potential in generative tasks by learning data distributions through iterative noise addition and denoising. Inspired by DiffusionDet, which formulates object detection as a reverse diffusion process, this paper proposes a novel multiscale SAR ship detection method based on a diffusion model. During training, our model learns to recover ground-truth boxes from noised versions via a Markov chain. At inference, it detects ships of various scales directly from random boxes without anchor-based priors. To mitigate false alarms on land, we further introduce a segmentation-derived land-sea mask as structural prior knowledge. Extensive experiments on SAR ship datasets demonstrate that our method achieves competitive accuracy and robustness across diverse marine environments, offering a new perspective for scalable and efficient SAR target detection.
