现代雷达2025，Vol.47Issue(11)：52-57,6.DOI:10.16592/j.cnki.1004-7859.2025092201
基于频域注意力Diffusion Transformer的SAR舰船图像生成技术
Frequency-domain Attention Enhanced Diffusion Transformer for SAR Ship Image Generation
摘要
Abstract
Currently,synthetic aperture radar(SAR)image generation methods based on generative adversarial networks are con-strained by the inherent instability during training,while diffusion model-based methods still primarily rely on the traditional U-Net backbone.A novel diffusion model architecture,termed spectral diffusion transformer—SpectDiT,is proposed in this paper,ai-ming to achieve high-quality SAR image generation.This model integrates spectral layers with transformer attention layers,and in-troduces frequency-domain features for modeling in the diffusion process,thereby further enhancing the quality and realism of SAR images.Compared to traditional denoising diffusion probabilistic models and the diffusion transformer model based on full attention Transformers,SpectDiT demonstrates superior performance in the task of SAR image generation,achieving new optimal perform-ance particularly in metrics such as peak signal-to-noise ratio,structural similarity index measure,and learned perceptual image patch similarity.Notably,SpectDiT possesses a flexible architecture design,where the ratio of spectral layers to attention layers can be adjusted to adapt to different generation tasks.As a new backbone network for diffusion models,SpectDiT breaks new ground in SAR image synthesis and holds the potential for extension to image generation tasks in other fields.关键词
去噪扩散概率模型/扩散Transformer/合成孔径雷达图像生成/频域学习Key words
denoising diffusion probabilistic model(DDPM)/diffusion transformer/synthetic aperture radar(SAR)image genera-tion/frequency-domain learning分类
信息技术与安全科学引用本文复制引用
黄文宇,熊刚,舒汀,郁文贤..基于频域注意力Diffusion Transformer的SAR舰船图像生成技术[J].现代雷达,2025,47(11):52-57,6.基金项目
上海自然科学基金资助项目(25ZR1401197) （25ZR1401197）
国家自然科学基金资助项目(62071293) （62071293）
