基于深度学习的机载高分辨率SAR图像运动舰船目标分类研究

With the strategic needs of the country and the military applications, it is very urgent to strengthen the rapid detection, accurate identification and accurate verification of ship targets,. Compared with space-borne SAR, airborne SAR has the advantages of flexibility in sensing ground and sea, and has unique characteristics in the field of ship target classification and recognition. In the classification and recognition of moving ship targets on airborne SAR images, there are some problems, such as blurred image defocusing caused by ship motion, weakened scattering characteristics caused by small ground angle of airborne SAR system, and fewer effective training sample sets when applying depth learning technology. This research is based on the strong feature extraction and classification ability of depth learning technology. In the aspect of sample pretreatment and network model improvement, optimization methods are put forward, which are based on ISAR fine focusing imaging technology of moving ship, and deep convolution neural network technology of introducing attention model to solve the scientific problem of difficult classification and recognition of moving ship in airborne SAR image, so as to improve the classification and recognition accuracy of moving ship targets. Through the experimental analysis of the airborne SAR data accumulated by the Academy of China Air Force over the years, the technical methods proposed in this project are verified.

面向国家”一带一路”战略需求和当前迫切的军事应用需求，加强舰船目标快速检测、准确识别和精确查证研究十分迫切。相比于星载SAR，机载SAR对地对海观测有着机动灵活的优势，在舰船目标分类识别领域也有着独特的特点。机载SAR图像上运动舰船目标分类识别时，存在着舰船运动引起成像散焦特征模糊、机载SAR系统小擦地角引起舰船上层建筑散射特征弱化，和应用深度学习技术时有效训练样本集少的问题，本研究基于深度学习技术强大的特征提取与分类能力，针对上述问题，分别在样本预处理和网络模型改善方面提出优化方法，即基于运动舰船ISAR精聚焦成像的预处理技术，和引入注意力模型构建深度卷积神经网络技术，解决机载SAR图像上运动舰船分类识别困难的科学问题，提升运动舰船目标分类识别精度。通过对课题依托单位历年积累的机载SAR数据实验分析研究，验证本课题提出的技术方法。

针对不同海况、不同大小舰船研究开发稳健的 ISAR 精聚焦处理.算法，并在其基础上开展基于深度学习的舰船目标检测、分类和识别.处理，研制 SAR 载荷舰船 ISAR 聚焦与分类识别软件。具备 SAR 幅.度图像舰船检测标记功能；具备基于 SAR 原始回波数据或者 SAR 原.始复图像数据开展舰船 ISAR 聚焦功能；具备基于 ISAR 图像切片开.展智能化目标分类识别功能。