摘要
在军事背景下,实现SAR图像的更快更精准分类能够取得战略优势。首先论述了军事SAR图像在分类的过程中,生成的图像特征偏向于多数类样本致使分类精度低的现状。然后,基于改进的DCGAN(生成对抗网络)的方法与数据增强理论,对少数类样本扩增。同时,将DCGAN中的判别器嵌入一个多分类器,使得在分类过程中不仅对图像真假分类,也能够对真实类别分类。最后,基于Kappa值与训练和测试精度,将所生成的结果与优化CNN、VGG、DLN、Resnet网络训练结果对比,发现通过改进的DCGAN生成的SAR图像可以提升分类精度,并且具有更好的泛化能力。
Under the military background, the realization of faster and more accurate classification of SAR images can gain strategic advantages. Firstly, during the classification process of military SAR images, the current situation that the generated image features are skewed to most samples and the classification accuracy is low are discussed. Then, based on the improved DCGAN(Generation Adversarial Network) method and data augmentation theory, a small number of samples were amplified. At the same time, the discriminator in DCGAN is embedded into a multi-classifier, which makes it possible to classify not only true and false images, but also real categories in the classification process. Finally, based on the Kappa value and training and test accuracy, comparing the generated results with the optimized CNN, VGG, DLN, and Resnet network training results. It is found that the SAR images generated by the improved DCGAN can improve the classification accuracy and have better generalization ability.
作者
鲁力
王洁
韩要昌
吴亚晖
LU Li;WANG Jie;HAN Yao-chang;WU Ya-hui(Graduate School,Air Force Engineering University,Xi'an 710051,China;Air Force Early Warning Academy Radar Officer School,Wuhan 430345,China)
出处
《控制工程》
CSCD
北大核心
2020年第3期561-566,共6页
Control Engineering of China
关键词
DCGAN
图像分类
数据增强
多分类器
DCGAN
image classification
data enhancement
multi-classifier
作者简介
通讯作者:鲁力(1996-),男,湖北荆州人,研究生,主要研究方向为SAR图像处理等。
