基于空间-分数谱域联合稀疏表示的SAR图像目标识别

Based on the theory of sparse approximation and compressive sensing,this project addresses the issues of feature extraction and automatic target recognition (ATR) for synthetic aperture radar（SAR）images. The main contributions are: To meet the challenge of the recognition of low-resolution, low-SNR targets, the fractional frequency analysis is introduced into the SAR feature extraction and recognition for the first time.To fuse the features in spatial and frequency domains and achieve an accurate and effective feature representations, a joint sparse representation method is proposed. The classification performances of exsiting sparse dictionaries for SAR ATR are not satisfactory due to the lack of discrimination and robustness to noise. To overcome this problem,the goals of providing discrimination,noise reduction as well as compression are corporated into the objective of dictionary training.A new training model is proposed to learn simultaneously reconstructive,discriminative and compressive as well as noise-robust dictionaries. Based on this,a joint sparse coding and classification method is proposed to improve the classification accuracy and robustness to environmental variations. This project focuses on the innovative theory and methods of modern signal processing and has a good prospect with significant scientific research achievements.

在稀疏逼近和压缩感知理论基础上，研究合成孔径雷达(Synthetic Aperture Radar，SAR)图像特征提取和自动目标识别的新方法。本项目的重要创新点：针对低分辨率、低信噪比的微弱SAR地面目标检测识别困难的问题，首次将分数谱分析引入到SAR的特征提取和目标识别中；用联合稀疏表示的方法将SAR空域与分数谱域特征进行融合，在挖掘其相关性的基础上提高表示的准确性和有效性；针对现有SAR稀疏字典不具备辨别性，且信号表示能力受噪声影响等问题，将提高辨别能力、去除噪声、可压缩等目标融合到字典训练中，得到一个集表示性、辨别性、噪声抑制性及可压缩性为一体的特征字典。在此基础上，提出联合稀疏分解和识别的方法，在提高识别准确率的同时增强对目标特性(如方位角、姿态)及噪声等因素变化的鲁棒性。本项目着眼于现代信号处理的新理论和方法研究，有望取得具有国际先进水平的学术和科研成果。

在稀疏逼近和压缩感知理论基础上，研究合成孔径雷达(Synthetic Aperture Radar，SAR)图像特征提取和自动目标识别的新方法。研究在图像空间域和分数谱域中SAR目标的特征提取方法，在挖掘两类特征相关性的基础上，找到特征的联合表示方法，并依据联合稀疏模型研究新的SAR目标识别方法。针对低分辨率、低信噪比的微弱SAR地面目标检测识别困难的问题，首次将分数谱分析引入到SAR的特征提取和目标识别中；用联合稀疏表示的方法将SAR空域与分数谱域特征进行融合，在挖掘其相关性的基础上提高表示的准确性和有效性。在此基础上，提出联合稀疏分解和识别的方法，在提高识别准确率的同时增强对目标特性(如方位角、姿态)及噪声等因素变化的鲁棒性。系统研究了SAR图像表示相关理论及方法、基于稀疏表示的SAR目标识别方法、分数域频谱成像理论及其关键技术、稀疏表示理论及其关键技术。利用分数域时频变换获得了信号的高分辨频谱特征，将其与图像空间像素域特征进行联合稀疏表示和稀疏分类，有效地改善了现有方法在特征提取和特征分类中的问题，提高了SAR目标识别的准确率。本项目着眼于现代信号处理的新理论和方法研究，取得了具有国际先进水平的学术和科研成果。