MeG-DiffSAR: Diffusion-Guided Synthesis of Realistic SAR Sea–Ship Scenes

The synthesis of synthetic aperture radar (SAR) images, particularly for mixed scenes, has significant potential in addressing data scarcity to support the training of deep-learning-based SAR models. In this letter, we present MeG-DiffSAR, a diffusion-based network for synthesizing SAR sea–ship scenes. First, we introduce a localized generation strategy that formulates high-fidelity mixed-scene synthesis as an inpainting task, focusing primarily on the boundary regions between ships and the surrounding sea surface. In order to enhance the realism of synthesized images, then a mask-enhanced multi-head self-attention mechanism (MeMSM) is proposed to guide the network’s focus toward masked regions, thereby generating contextually consistent inpainting results. At last, a gradient-aware loss is incorporated to improve the textural fidelity and visual quality of synthesized images. The experimental results demonstrate the effectiveness of our method. Especially, compared to the state-of-the-art method Ship-Go, MeG-DiffSAR reduces FID by 13.64 and improves PSNR by 14.17dB.