大幅SAR图像的复杂背景下多目标认知技术的研究

Due to multiple-targets and changing backgrounds in large-scale radar SAR images, extracting targets involved in complex background is a difficult field of study in SAR target detection and interpretation. It aspires to achieve cognitive Radar targets by combining analysis of Statistical properties, target detection principles and SAR image interpretation based on a more comprehensive database of measured targets and complex backgrounds. At first, the algorithm for multi-target detection is researched based on Statistical Properties and Compressed Sensing Technology, which is divided into two parts: (1) How to search fields for the captured objects by using statistical diagnostics is researched in different clutter backgrounds and objectives. Then, the algorithms of intelligent automatic screening CFAR target detection are explored in complex clutter backgrounds. (2) The algorithm for SAR image multi-target detection based on CS is proposed. In this algorithm, the detection results of different targets in the same SAR image are considered as image information of the same scene which is detectable in different modes. Additionally, pre-processing, feature extraction and recognition, and the interpretation of algorithms for SAR image targets are studied.

由于大幅雷达SAR 图像中目标种类多样，目标数量较多，并且所处的背景多变，所以如何在复杂背景下确定所需目标地域、提取所需目标成为目前雷达目标检测研究的难点。本项目在目标和复杂背景实测数据库的基础上，将SAR图像统计特性分析、目标检测理论及目标解译相结合，达到对雷达目标完全认知的目的。首先研究基于统计诊断和压缩感知技术的大幅SAR 图像目标检测算法：将SAR 图像的复杂背景和目标通过数据挖掘中的统计诊断分析来确定所需目标区域从而构建SAR 图像各种背景情况下的智能自动筛选CFAR 多目标检测算法；再将不同模式下获取检测图像信息通过压缩感知技术融合得到更为可靠的检测目标。其次，针对以上算法检测出的多种目标进行解译研究，包括预处理、特征提取以及识别算法等。

由于大幅雷达SAR 图像中目标种类多样，目标数量较多，并且所处的背景多变，所以如何在复杂背景下确定所需目标地域、提取所需目标成为目前雷达目标检测研究的难点。本项目在目标和复杂背景实测数据库的基础上，将SAR图像统计特性分析、目标检测理论及目标解译相结合，达到对雷达目标完全认知的目的。本项目提高了SAR对不同背景、不同目标的认知能力；研究了多目标智能自动检测中对不同背景的统计特征提取，辨别背景缩小侦查范围，检测的准确率达到90%；研究了多目标智能自动检测中CFAR自动智能筛选检测器中关键技术和参数的设置，适应多种不同目标检测任务，在信噪比大于5dB时检测率达到80%，虚警概率小于0.1；研究了基于压缩感知技术对于图像融合的相关测度矩阵的新概念及求解方法，完成多幅检测目标图像的融合，利用CS的图像融合技术才能更精准的检测目标，提高检测率和目标完整性，也为多传感器图像融合目标检测提供新的解决途径，使获得的目标更为清晰可靠，检测率提高10个百分点以上；由于任何单一的分类器都不能完全解决面临的问题或者达到系统的要求，所以在此需要研究SAR解译中相干斑预处理的MRF复原模型，并对目标的边缘、散射中心、峰值、统计特性等特征提取，来实现多目标间的种类及型号多特征融合的识别，完成SAR图像目标解译识别整个过程，种类识别达到90%，型号识别达到80%以上，并搭建一套完整的SAR目标认知指标体系，给解译和识别程度提供指导标准。