Ship-SAR: an improved ship image generation framework of synthetic aperture radar via denoising diffusion probabilistic model

Synthetic Aperture Radar (SAR) has become a highly regarded technique in the field of remote sensing due to its all-weather and all-time imaging capabilities. To address the limited generalization of SAR ship detection models caused by scarce public remote sensing data and restricted scenarios, this paper proposes an improved high-quality image generation framework Ship-SAR based on Denoising Diffusion Probabilistic Models (DDPM). Firstly, the denoising network UNet extracts the long-tail contours and multi-scale texture features of ships step by step by combining a reparameterization multi-branch structure with an efficient layer aggregation network. Secondly, this paper proposes a lightweight adaptive downsampling module that mitigates information loss in traditional downsampling processes by retaining task-relevant semantic information through a dynamic feature selection mechanism. Additionally, a depth spatial upsampling block is introduced to rearrange channel dimension information into the spatial dimension, which achieves resolution enhancement. Thirdly, Ship-SAR introduces an epsilon scaling strategy to mitigate input data bias between training and sampling stages. This approach corrects sampled trajectories without requiring retraining, significantly enhancing generation quality. Ship-SAR generates SAR ship images that compare favorably to baseline DDPM in terms of Fréchet inception distance metrics and structural similarity index measure. YOLOv11s achieved an increase in mAP from 87.4% to 90.9% on high-resolution SAR image datasets and an increase in mAP of 2.1% for RTDETR-r50. The comparative results show that the SAR ship images generated by Ship SAR significantly improve the diversity of the Ship-SAR image dataset and promote progress in the field of remote sensing detection.