基于字典学习和测量矩阵优化的SAR图像目标识别方法研究

Traditional Synthetic Aperture Radar (SAR) target recognition method is limited by feature extraction and computing complexity. Topic proposed in the framework of compressed sensing theory to SAR sparse signal compression measurement and target recognition. Study on the sparse characteristics of SAR images in different transformation dictionary, sparse matrix modeling, including dimensional compression adaptive over-complete dictionary for image degradation, adaptive over-complete dictionary for image sparse nearest neighbor representation, Compressive sampling matching pursuit method, nonparametric Bayesian adaptive sparse decomposition method, measurement matrix optimization modeling, structured random matrix optimization design method, the uncertainty measurement matrix optimization design method, and the experiment analysis of SAR ATR. Study aims to overcome the radar target recognition method of feature extraction sensitivity, and form a SAR ATR theory and method based on adaptive compressed sensing. Using optimization theory and modern intelligent information processing technology, we develop the SAR image dimension reduction, sparse representation method, the fast convergence of the adaptive sparse decomposition algorithm, and the measurement matrix optimization design method. Based on a dictionary learning adaptive over-complete dictionary for the optimization design of the measuring matrix of image sparse representation can be implemented, and the sparse decomposition can be used to SAR image target recognition. Research results could lead to the development of a new type of radar target recognition system and the key to improve technical support.

传统合成孔径雷达目标识别方法受限于特征提取和计算度复杂。本课题拟在压缩感知理论框架下，以SAR稀疏信号压缩测量及目标分类识别为主线，研究SAR图像在不同变换字典下的稀疏特性、图像降维压缩自适应过完备字典的构造、图像稀疏近邻表示字典构造、压缩采样匹配追踪方法、非参数贝叶斯自适应稀疏分解方法、测量矩阵优化模型、随机测量矩阵优化设计方法、确定性测量矩阵的优化设计方法，SAR ATR的实验验证分析。研究旨在克服雷达目标识别方法对特征提取敏感性，形成基于自适应压缩感知SAR ATR理论与方法。利用优化理论、现代智能信息处理技术，发展SAR图像降维、稀疏表示方法和快速收敛的自适应稀疏求解算法，以及测量矩阵优化设计方法。基于字典学习的自适应过完备字典构造和测量矩阵的优化设计能够实现对图像稀疏表示，并在稀疏分解后用于目标分类识别。研究成果有望形成对于我国新型雷达目标识别系统的研制和改进的关键技术支撑。

基于自适应压缩感知合成孔径雷达图像目标识别是当前研究热点，引起学术界关注。本项目组开展的基于字典学习和测量矩阵优化的SAR图像目标识别研究，重点突破了SAR图像在自适应过完备字典下的稀疏表示、快速稀疏求解方法和测量矩阵优化等关键技术。具体包括：1）建立了SAR图像稀疏模型，提出了自适应过完备字典的构造方法；2）提出了非参数贝叶斯自适应快速稀疏求解方法；3）提出了随机测量矩阵和确定性测量矩阵的优化设计方法；4）开展了MSTAR数据库目标的实验验证。研究成果为后续目标识别、跟踪提供了基础，部分成果应用在空间监视等需求背景中。. 项目组依托研究成果，发表高水平学术论文19余篇，其中在IEEE/IET期刊发表文章16 篇，SCI检索4 篇，EI检索12 篇。