| 摘要: |
| 针对雷达辐射源数据集样本数量有限、样本多样性不足的问题，提出了一种无监督的由图像生成图像的网络—多样性最大化生成对抗网络(Maximum Diversity Generative Adversarial Network，MDGAN)。该网络在原始生成对抗网络的生成器目标函数基础上加上了一个额外的正则化项，该正则化项表示生成器中特征图之间的距离与生成特征图所用随机向量之间的距离的比值，通过最大化这个比值，可以让生成器尽量生成拥有不同特征的样本，以增加样本的多样性。对6种常见雷达信号进行仿真实验，证明了MDGAN在生成真实且多样的样本方面是有效的。 |
| 关键词: 雷达辐射源识别 多样性最大化生成对抗网络 样本多样性 起始分值 弗雷歇起始距离 |
| DOI：DOI:10.3969/j.issn.1672-2337.2020.02.016 |
| 分类号:TN971 |
| 基金项目:复杂电磁环境效应国家重点实验室项目（No.2018Z0202B） |
|
| Data Set Expansion with Maximum Diversity Generative Adversarial Network |
|
LI Kun，ZHU Weigang
|
|
Space Engineering University，Beijing101416，China
|
| Abstract: |
| To deal with the problem of limited sample size and insufficient sample diversity for radar emitter data sets，a maximum diversity generative adversarial network (MDGAN) is proposed using an unsupervised image to image network.The network adds an additional regularization term to the generators objective function of the original generative adversarial network. The regularization term represents the ratio of the distance between the feature maps in the generator and the distance between the random vectors used to generate the feature maps. By maximizing this ratio，the generator can generate samples with different features as much as possible to increase the sample diversity.Experiments with six common radar signals demonstrate that MDGAN is effective in generating real and diverse samples. |
| Key words: radar emitter identification maximum diversity generative adversarial network (MDGAN) sample diversity inception score(IS) Frechet inception distance(FID) |
