基于多尺度GAN 网络的SAR 舰船目标扩充 |
| |
引用本文: | 黄琼男.基于多尺度GAN 网络的SAR 舰船目标扩充[J].兵工自动化,2022,41(7). |
| |
作者姓名: | 黄琼男 |
| |
作者单位: | 航天工程大学电子与光学工程系 |
| |
摘 要: | 针对构建合成孔径雷达(synthetic aperture radar,SAR)图像舰船目标数据集的过程中,某些舰船类型样本
不足的问题,提出一种基于多尺度生成对抗网络(IC-ConsinGAN)的SAR 舰船目标扩充方法。通过将注意力机制引
入并行多阶段多尺度GAN 网络中,提取SAR 舰船目标的关键特征,抑制背景特征,使得生成的SAR 图像舰船目标
不仅具有精细化结构,而且弥补了单幅图像生成过程中多样性不足的问题。实验结果表明:SIFID 指标比原始
ConsinGAN 网络模型下降了0.02,将扩充数据加入到SAR 舰船目标识别任务中,10 类舰船目标平均识别率提升了
8.4%,证实了IC-ConsinGAN 模型的有效性,具有一定的工程应用价值。
|
关 键 词: | 生成对抗网络 合成孔径雷达 注意力机制 多尺度 舰船目标识别 |
收稿时间: | 2022/3/20 0:00:00 |
修稿时间: | 2022/4/24 0:00:00 |
SAR Ship Target Expansion Based on Multiscale GAN Network |
| |
Abstract: | In order to solve the problem of insufficient samples of some ship types in the process of constructing
synthetic aperture radar (SAR) image ship target data set, a SAR ship target expansion method based on multiscale
generated countermeasure network (IC-ConsinGan) is proposed. By introducing the attention mechanism into the parallel
multi-stage multiscale GAN network, the key features of SAR ship targets are extracted and the background features are
suppressed, so that the generated SAR image ship targets not only have a refined structure, but also make up for the lack of
diversity in the process of generating a single image. The experimental results show that the SIFID index is 0. 02 lower than
that of the original ConsinGan network model, and the average recognition rate of 10 types of ship targets is improved by
8.4% when the extended data is added to the SAR ship target recognition task, which confirms the effectiveness of the
IC-ConsinGan model and has certain engineering application value. |
| |
Keywords: | generative countermeasure network synthetic aperture radar attention mechanism multiscale ship
target recognition |
|
| 点击此处可从《兵工自动化》浏览原始摘要信息 |
|
点击此处可从《兵工自动化》下载全文 |