基于CDCGAN的SAR图像数据增广  

SAR Image Augmentation Method Based on CDCGAN

作　　者：赵竹新 范纯卓 刘艳博[1] 冯彦卿 王海强 ZHAO Zhuxin;FAN Chunzhuo;LIU Yanbo;FENG Yanqing;WANG Haiqiang(Beijing Institute of Remote Sensing Information,Beijing 100011,China)

机构地区：[1]北京市遥感信息研究所,北京100011

出　　处：《无线电工程》2025年第3期580-587,共8页Radio Engineering

摘　　要：针对合成孔径雷达(Synthetic Aperture Radar,SAR)图像采集成本高、多样性不足,影响图像解译效果的问题,在现有深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Network,DCGAN)的基础上,提出了基于条件输入的DCGAN模型,实现了方位角/俯仰角/斜视角可控的SAR图像增广,完成了对现有仿真数据集的扩充。建立了SAR增广图像评价指标体系,对增广图像进行了客观的质量评估。结果表明,提出的SAR图像增广方法可以高质量地实现SAR图像多角度可控样本扩增,对于提升SAR图像角度丰富性具有积极意义。To address the problems of high cost and lack of diversity of Synthetic Aperture Radar(SAR)image acquisition which affect the effect of image interpretation,on the basis of existing Deep Convolutional Generative Adversarial Network(DCGAN),this paper proposed a model based on conditional DCGAN(DCGAN with conditional input)which realizes SAR image augmentation with customized azimuth/pitch/oblique angle and greatly expands the existing simulation dataset.Also,an evaluation index system of SAR augmented images is established to evaluate the quality of SAR augmented images objectively.The results show that the proposed method can successfully achieve the expansion of SAR image samples with multiple customized angles with satisfied quality and lower time cost,which is of positive significance for enhancing the diversity of SAR image angles.

关 键 词：合成孔径雷达仿真 语义可控条件生成对抗网络 卷积 角度信息可控 

分 类 号：TP752[自动化与计算机技术—检测技术与自动化装置]