基于注意力深度网络的SAR目标检测识别一体化方法

SAR target detection and target recognition are two very important parts in SAR image information interpretation technology. Each of them has relatively independent and perfect theory and method system. However, from the task requirement of SAR system equipment, especially for the target recognition problem in large scale scene SAR image, the target detection and target recognition should be combined. So it is urgent to establish the SAR target detection and recognition integration method, in order to to meet the equipment application requirement of SAR target recognition method..The method of SAR Target Detection and Recognition Integration via Attention Deep Networks is proposed in this project, which can directly and fast recognize the targets in high resolution and large scale scene SAR images. At the bases of fusing the traditional detection and recognition process in one deep network framework, this method can make full use of spatial information between target and background in large scene SAR images, and the feature information of the targets, via the spatial attention mechanism and features attention mechanism. .This method is mainly focuses on solving two key scientific problems: the saliency target region regression based on spatial attention and the saliency target features selection bsaed on feature attention, which can break through the technology bottleneck of traditional End-to-End recognition mechanism, recognize the SAR targets from high resolution large scene images directly, and can improve the intelligent and practical level of SAR target recognition.

SAR目标检测与目标识别是SAR图像信息解译技术的两个重要组成部分，各自都有着较为独立和完善的理论与方法体系。而从SAR系统装备的任务需求出发，针对大场景SAR图像中的目标识别问题，需要目标检测与目标识别方法有机结合，建立SAR目标检测识别一体化方法，以满足SAR目标识别方法的装备应用需求。.本项目提出基于注意力深度网路的SAR目标检测识别一体化方法，在将目标检测和目标识别有机融合在一个深度网络框架的基础上，通过空间注意力机制和特征注意力机制，充分利用大场景SAR图像中目标与背景的空间信息以及目标本身的特征信息，能够快速实现高分辨大场景SAR图像中目标的直接识别。.该方法主要解决基于空间注意力的显著性目标区域回归，以及基于特征注意力的显著性目标特征优选等关键科学问题，能够突破传统End-to-End识别机制模型的技术瓶颈，提升SAR目标识别方法的智能化和实用化水平。

为实现大场景SAR图像中目标的快速发现、精确定位、准确识别，满足SAR目标识别方法的装备应用需求，需要有机结合目标检测与目标识别环节，建立目标检测识别一体化方法，突破目标检测识别串行处理方式的局限，从识别方法和识别框架上解决传统End-to-End识别机制存在的问题，提升SAR目标识别方法的应用能力。.项目在研究过程中提出了基于深度网络的SAR目标检测识别一体化方法、基于密集注意力金字塔网络的SAR图像中多尺度舰船目标检测、以及基于空间注意力的大场景SAR图像中舰船目标检测等方法，能够实现大场景SAR图像中车辆、舰船等目标的快速识别。针对约7千万像素点的实测SAR数据，本项目提出的方法可以达到90%以上的正确检测率与识别率，且处理时间小于32秒。.项目总体按照计划执行，达到了预期目标，也实现了预期成果。该项目部分算法已应用于北京华航无线电测量研究所，用于某型号装备针对机场中飞机目标检测的试验验证；也应用于上海无线电设备研究所，为其某平台实现舰船目标检测提供技术支撑。