| 摘要: |
| 针对临近空间高速目标识别任务中存在的高分辨一维距离像（High Resolution Range Profile, HRRP）数据样本稀缺、噪声干扰严重等问题，本文提出一种基于稀疏去噪变分自编码器（Sparse Denoising Variational Autoencoder，SDVAE）提取的深度抽象特征与极化旋转域物理特征融合的临近空间高速目标识别方法。首先，在特征提取阶段，通过设计具备稀疏约束的SDVAE模型提取全极化HRRP数据的深度特征，即在编码器中注入高斯噪声以模拟真实环境，并在潜在空间中引入稀疏惩罚项筛选关键特征，从而在小样本和低信噪比条件下获得更好的泛化能力。其次，对全极化HRRP引入极化旋转域特征，并从中提取十项具有鲁棒性、物理可解释性的极化参数。然后，利用注意力机制实现深度特征与极化旋转域特征的自适应融合，并通过多层感知机完成分类识别。最后，基于临近空间高速目标动态全极化HRRP数据集的实验结果验证了本文方法的有效性，以及在小样本、低信噪比等条件下相比传统方法的优势。 |
| 关键词: 临近空间高速目标 高分辨距离像 全极化 变分自编码器 极化旋转域 |
| DOI： |
| 分类号:TN957 |
| 基金项目:国家自然科学基金(62401589),湖北省自然科学基金(2024AFB653),湖北省教育厅科学研究计划重点项目(D20241503) |
|
| A Full-Polarimetric One-Dimensional Range Profile–Based Method for Near-Space High-speed Space Target Recognition |
|
|
|
|
| Abstract: |
| To address the challenges of data scarcity and severe noise interference in near-space high-speed target recognition using High-Resolution Range Profiles (HRRP), this paper proposes a novel recognition method. This approach fuses deep abstract features extracted via a Sparse Denoising Variational Autoencoder (SDVAE) with physical features derived from the polarization rotation domain. First, in the feature extraction phase, an SDVAE model incorporating sparse constraints is designed to extract deep features from fully polarimetric HRRP data. Specifically, Gaussian noise is injected into the encoder to simulate realistic environmental conditions, while a sparsity penalty term is introduced in the latent space to screen for salient features, thereby enhancing generalization capabilities under conditions of limited samples and low signal-to-noise ratios (SNRs). Subsequently, the polarization rotation domain is investigated for fully polarimetric HRRP, from which ten robust and physically interpretable polarization parameters are extracted. An attention mechanism is then utilized to achieve the adaptive fusion of the deep features and the polarization rotation domain features, followed by target classification via a Multi-Layer Perceptron (MLP). Finally, experimental results based on a dynamic fully polarimetric HRRP dataset of near-space high-speed targets demonstrate the effectiveness of the proposed method, confirming its superiority over traditional approaches, particularly under challenging conditions involving small sample sizes and low SNRs. |
| Key words: Near-space high-speed space targets High resolution range profile Full-polarization Variational autoencoder Polarization rotation domain. |