基于深度学习的逆合成孔径雷达成像技术研究

Intelligent radar is a growing development trend. This project intends to carry out the research of inverse synthetic aperture radar (ISAR) imaging technology within the framework of deep learning (DL) theory. The powerful learning ability of the deep neural network (DNN) is exploited to extract the abstract structural information in the ISAR data. New ideas and new methods exploiting the advantages of DNN are expected to appear to enhance the level of ISAR imaging technology and break through the bottlenecks of the existing ISAR imaging technology. Firstly, DNN-based ISAR imaging methods with different architectures are proposed. The feasibility, implementation and advantages associated with each DNN architecture are analyzed. A DNN-based forward model for ISAR imaging is built and then is utilized to better tackle the imaging problems of non-cooperative targets. ISAR motion compensation exploiting the DL technology is studied to overcome the limitation of exiting motion compensation technologies. The construction and enhancement the ISAR training data sets are studied for completing the aforementioned works. The comparison of the proposed methods with the traditional ISAR imaging methods and the compressive sensing based imaging methods is performed to clarify the advantages of DL-based ISAR imaging. The research is expected to provide theoretical support for intelligent radar imaging technology.

瞄准雷达智能化发展趋势，本项目拟在深度学习(DL)理论框架下开展逆合成孔径雷达(ISAR)成像技术研究。利用深度神经网络(DNN)的强大学习能力，提取目标ISAR数据中的抽象结构信息来成像，为提升ISAR成像技术水平提供新思路和新方法，突破现有ISAR成像技术瓶颈。提出基于深度神经网络(DNN)的ISAR成像方法，明确不同网络架构实现ISAR成像的可行性、实现方法和优势；研究基于DNN的ISAR成像前向模型构建并发挥其在非合作目标成像中的作用；研究基于DNN的ISAR运动补偿，突破现有运动补偿技术的局限性；配合上述研究，提供ISAR训练数据集构建和增强方法。通过与传统ISAR成像方法和压缩感知成像方法的对比，明确DL应用于ISAR成像中的优势所在，为智能雷达成像技术提供理论支持。

本项目以智能化逆合成孔径雷达(ISAR)成像技术发展为牵引，在DL框架下，研究利用深度神经网络(DNN)，挖掘ISAR数据的深层特征信息，解决雷达成像问题，提出全新的雷达成像方法。分别构建了CNN、FCNN、G-FCNN、TL-G-FCNN和CV-CNN，明确了各种网络架构的可行性、实现方法和优势。研究了基于深度ADMM网络的ISAR成像方法，明确了模型驱动网络和数据驱动网络用于ISAR成像的不同。配合上述研究，研究了基于实测数据的训练样本增强方法和基于电磁仿真数据的训练数据集构建方法。研究基于FCNN的ISAR自聚焦方法，利用深度特征信息实现ISAR数据自聚焦。通过与传统ISAR成像方法和压缩感知成像方法的对比，明确DL应用于ISAR成像中的优势所在，为智能雷达成像技术提供理论支持。此外，在该项目研究过程中，并研究外拓至干涉ISAR三维成像方法和SAR/ ISAR目标识别方法。