作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (9): 259-265. doi: 10.19678/j.issn.1000-3428.0061223

• 图形图像处理 • 上一篇    下一篇

基于EMD-MDGAN的HRRP增扩方法

王紫娇, 王晓丹   

  1. 空军工程大学 防空反导学院, 西安 710051
  • 收稿日期:2021-03-21 修回日期:2021-05-07 发布日期:2021-09-13
  • 作者简介:王紫娇(1996-),女,硕士研究生,主研方向为智能信息处理;王晓丹(通信作者),教授、博士、博士生导师。
  • 基金资助:
    国家自然科学基金(61876189)。

HRRP Expansion Method Based on EMD-MDGAN

WANG Zijiao, WANG Xiaodan   

  1. Air and Missile Defense College, Air Force Engineering University, Xi'an 710051, China
  • Received:2021-03-21 Revised:2021-05-07 Published:2021-09-13

摘要: 高分辨率距离像(HRRP)是弹道目标识别的主要特征,由于其为非合作目标因此观测频率极低,导致带标签样本量严重不足,而混合密度生成对抗网络(MDGAN)作为生成HRRP的有效方法,存在模式崩溃、网络不易收敛等问题。提出一种基于误差匹配分布(EMD)改进MDGAN的弹道目标HRRP增扩方法EMD-MDGAN。将生成器、残差网络和注意力机制相结合,通过残差结构解决梯度消失的问题,利用注意力机制提高生成器中自编码器的特征提取能力,并把误差匹配思想引入损失函数设计中,以增强模型的稳定性,使网络更易收敛。实验结果表明,该模型在有效解决模式崩溃问题的基础上,可缩小生成样本与真实样本分布间差异,生成具有一定真实性、可靠性、多样性的数据,实现HRRP数据增扩。

关键词: 混合密度生成对抗网络, 残差网络, 注意力机制, 误差匹配分布, 高分辨率距离像

Abstract: High-Resolution Response Profile(HRRP) is one of the main features for ballistic target recognition, but it is a non-cooperative target and thus limited by the low observation frequency, leading to a lack of labeled samples. Mixture Density Generative Adversarial Network(MDGAN) is an effective way to generate HRRP, but deeply suffers from mode collapse and difficulties in network convergence.To address the problems, a new HRRP expansion method named EMD-MDGAN is proposed for trajectory targets.EMD-MDGAN is based on Mixture Density Generative Adversarial Network(MDGAN), which is improved by Error Matching Distribution(EMD).The method employs the residual network structure to solve the problem of gradient disappearance, and uses the attention mechanism to improve the feature extraction ability of the auto-encoder in the generator.On this basis, the idea of error matching is introduced into the design of loss function to enhance the stability of the model and make the network more convergent.The experimental results show that the model can avoid mode collapse, and thus reduce the distribution difference between the generated samples and actual samples.It can ensure the authenticity, reliability and diversity of generated data, and realize HRRP data expansion.

Key words: Mixture Density Generative Adversarial Network(MDGAN), residual network, attention mechanism, error matching distribution, High-Resolusion Response Profile(HRRP)

中图分类号: