基于深度学习的多源异质图像融合舰船识别方法研究

Ship classification and recognition is one of the key processing stages of the maritime monitoring system. It plays a significant role in the field of maritime traffic supervision，target reconnaissance，etc. Synthetic aperture radar (SAR) and optical imaging are two important means of ocean target monitoring, which have advantages and complement each other. However, there are essential differences between SAR and optical imaging mechanism. This factor makes it difficult to effectively fuse ship features from this two kinds of images, together with the differences in observation angle, resolution of the both sensors and the strong mobility of ships. Currently, this research is still not deep and thorough, due to the lack of multi-source datasets and the limited interpretation ability of SAR image. In this project, a standardized optical and microwave image datasets of typical ships will be established. On this basis, the imaging mechanism and characteristics differences between visible and SAR images will be studied and analyzed in depth. Then, we’ll constructe and optimize deep learning models for ship feature extraction, correlation and fusion of heterogeneous images, and further reveal the inherent correlation law of multi-source target features, in order to form a more perfect and accurate target feature representation. Furthermore, an automatic recognition algorithm is studied to availably solve the problems of generalization and slow training under the small-sample condition. Also, this task will evaluate and verify the recognition performances in many ways using measured data, so that the accuracy and robustness of recognition is improved. Consequently, this research can provide technical means and target characteristics support for intelligent monitoring of ocean targets. In addition, it can be flexibly extended to other target recognition applications by fusing infrared images as well as more various sensor data.

舰船分类识别在海运交通监管、海上目标侦察等领域具有重要的应用价值。合成孔径雷达（SAR）、光学成像作为海洋目标监测的两大重要手段，两者互有优势、互为补充。但是，SAR与光学成像机理具有本质区别；加之，两者成像视角、分辨率的不同以及海面舰船机动性强等，使得有效融合多源舰船特征存在诸多困难。受数据源以及SAR图像解译水平的限制，目前相关研究还处于初级阶段。本项目将在建立规范化的典型舰船光学、微波图像数据集的基础上，通过深入分析可见光与SAR图像的成像机理与特征差异，构建面向异质图像舰船特征提取与关联融合的深度学习优化模型，揭示多源特征的内在关联规律，形成更完善、确切的目标特征表示；建立小样本下泛化能力强、训练速度快的识别模型，并基于实测数据开展多方位性识别性能评估验证，提高识别精度与稳健性，为海洋目标智能监测提供技术手段与目标特性支持。此外，本研究可扩展于红外等更多源数据融合的目标识别应用。

舰船目标分类识别是海面目标监视系统的重要组成部分，在海运交通监管、海上目标侦察等领域具有重要的应用价值。合成孔径雷达（SAR）、光学成像作为海洋目标监测的两大重要手段，两者互有优势、互为补充。但是，SAR与光学成像机理具有本质区别；加之，两者成像视角、分辨率的不同以及海面舰船机动性强、海杂波与近岸环境的干扰等，使得有效融合多源舰船特征实现精细目标识别存在诸多困难。本项目提出了基于目标特性模型、生成对抗网络的海面舰船微波、光学图像生成方法；结合理论仿真、模拟测量、遥感实测手段，构建了高分辨率光学、微波图像典型军民舰船目标精细识别数据集；基于光电成像机理与特征分析方法，分析揭示了微波与光学图像中目标特征差异与内在关联规律，突破了基于异构深度卷积网络-高斯分布受限玻尔兹曼机的多源图像舰船特征提取与关联融合、识别等关键技术，形成了更完善、确切的目标特征描述模型，光电物理属性特征辅助的多源异质图像特征融合小样本舰船目标识别模型，实现了200幅量级样本规模，6类典型军民舰船识别率优于90%，相关成果可为海洋目标智能监测提供技术手段与目标特性支持。此外，所建立的多源特征融合模型具有轻量化，特征维度、网络结构扩展灵活的特点，可扩展应用于红外等更多传感器数据融合的海面、地面和空中目标识别。