航空兵器2025，Vol.32Issue(5)：114-120,7.DOI:10.12132/ISSN.1673-5048.2025.0052
对抗蒸馏扩散模型SAR图像相干斑抑制方法
Adversarial Distillation Diffusion for SAR Despeckling
摘要
Abstract
Speckle noise in synthetic aperture radar(SAR)images arises from the imaging mechanism and ap-pears granular,significantly interfering with image interpretation and the extraction of depth information.In this paper,two key aspects are investigated:model architecture and training data.Firstly,a SAR image denoising method based on adversarial distillation is proposed,termed AD-DPM,which integrates generative adversarial networks(GANs)with diffusion probabilistic models.Building upon the pre-trained conditional diffusion probabilistic model for denoising,de-noising diffusion implicit models(DDIM)are employed to accelerate model iteration generation.Furthermore,adver-sarial distillation is utilized to achieve a secondary reconstruction of images in the latent space.Secondly,a specialized dataset training model is developed by introducing multiplicative noise into multi-temporal denoised images.This ap-proach effectively circumvents structural and algorithmic errors inherent in optically simulated SAR images and over-comes limitations such as fixed noise intensity in multi-temporal denoising datasets.Experimental results demonstrate that AD-DPM achieves a 213-fold acceleration in conditional diffusion model generation,exhibits superior speckle sup-pression and preservation of original image structures,and displays good model generalization capabilities in random SAR image denoising tasks.关键词
SAR图像/相干斑噪声/扩散概率模型/对抗蒸馏/乘性噪声Key words
SAR image/speckle noise/DDPM/adversarial distillation/multiplicative noise分类
武器工业引用本文复制引用
贠相亚,左晓桐,谢志刚,王长龙,赵月飞..对抗蒸馏扩散模型SAR图像相干斑抑制方法[J].航空兵器,2025,32(5):114-120,7.基金项目
基础加强计划技术领域基金项目(2019-JCJQ-JJ-05) （2019-JCJQ-JJ-05）
