| 摘要: |
| 近场高分辨率合成孔径雷达（Synthetic Aperture Radar, SAR）成像通常需要采集大量数据，这增加了系统的复杂性和设备成本。为解决上述问题，已有相关研究将卷积神经网络应用于稀疏SAR成像领域，有效降低了数据量。然而，这类方法依赖于从大量数据中学习先验知识，其性能受到训练样本规模和质量的限制。为此，本文提出了一种基于无训练神经网络的SAR成像方法，无需额外的训练数据即可实现高分辨率SAR成像。首先，针对传统观测矩阵规模大存储困难的问题，引入近似观测算子以提升计算效率；其次，设计了一种硬件兼容的稀疏扫描策略，通过适配雷达实际扫描系统以减少数据量。最后，提出基于U-Net架构的CVUIU-Net（Complex-Valued U-Net in U-Net）模型，通过跨尺度特征融合增强局部细节与全局结构的表征能力，有效提升了边缘保持能力并抑制了成像伪影。仿真与实测实验结果表明，所提方法在重建质量上优于传统方法及现有神经网络基线模型，且在采样率降至10%、目标存在遮挡的场景下，仍能有效恢复目标信息，展现出较强的抗干扰能力和鲁棒性。 |
| 关键词: 近场 合成孔径雷达 稀疏成像 神经网络 |
| DOI： |
| 分类号:TN957 |
| 基金项目:上海市自然科学基金(No.24ZR1421800) |
|
| Near-Field SAR Sparse Imaging Based on Unsupervised Deep Learning |
|
|
|
|
| Abstract: |
| Near-field high-resolution Synthetic Aperture Radar (SAR) imaging typically requires the acquisition of large amounts of data, which increases system complexity and equipment costs. To address these challenges, previous studies have applied Convolutional Neural Networks (CNNs) in sparse SAR imaging, effectively reducing the data requirements. However, these methods rely on learning prior knowledge from large datasets, and their performance is limited by the scale and quality of the training samples. To overcome this, this paper proposes a SAR imaging method based on an untrained neural network that achieves high-resolution SAR imaging without the need for additional training data. First, to tackle the issue of large-scale storage difficulties associated with traditional observation matrices, an approximate observation operator is introduced to enhance computational efficiency. Second, a hardware-compatible sparse scanning strategy is designed, adapting to the constraints of the actual radar scanning system to reduce the amount of data acquisition. Finally, a Complex-Valued U-Net in U-Net (CVUIU-Net) model based on the U-Net architecture is proposed. This model improves the representation of local details and global structures through cross-scale feature fusion, effectively enhancing edge preservation and suppressing imaging artifacts. Simulation and experimental results demonstrate that the proposed method outperforms traditional approaches and existing neural network baseline models in terms of reconstruction quality. Even with a reduced sampling rate of 10% and in scenarios with target occlusion, the method effectively recovers target information, showcasing strong anti-interference capabilities and robustness. |
| Key words: near-field synthetic aperture radar sparse imaging neural network |