基于多任务拓扑学习的实时无人机航拍图像识别研究

As a newly emerging aerial platform, unmanned aerial vehicle (UAV) has attracted many attentions due to its high flexibility and low cost. It has wide range of applications in both civilian and military fields, such as low-altitude photogrammetry, marine monitoring, military reconnaissance and so on. UAV aerial images have the following features: 1) high resolution and large scale of dataset; 2) wide range of aerial photography, which includes numerous object categories and candidate recognition targets. These features make the real-time image categorization for UAV aerial image become a challenging and great important research area. On the basis of computer vision and machine learning theory and technique, this proposal will focus on two scientific issues: graph-based representation and matching for the content of aerial images, and topological selection and mining. Specific research tasks include: obtaining the knowledge of topological relationship between objects by construction of graph model for aerial images, integrating multiple topological learning and optimizing, and implementing the categorization model based on multi-task boosting algorithm. The research outcome is expected to provide fundamental theory and key technique support for video surveillance and scene annotation and will be expected to enhance the role of UAV aerial image categorization research in low-altitude remote sensing measurement and monitoring.

作为一种新兴的航空测量平台，无人机以其高灵活性、低成本的优势得到了广泛的关注，并被大量地应用于民用和军事领域，如低空摄影、海洋监测和军事侦察等。由于无人机航拍图像具有1）分辨率高、数据量大；2）区域广，包含地物目标类别丰富、待识别对象多等特点，使得针对无人机航拍图像的实时图像分类识别成为一个具有挑战性和重要意义的研究方向。本项目在计算机视觉、机器学习理论和技术的基础上，针对航拍图像内容的图模型表示与匹配、拓扑结构选择与学习两个关键问题，研究快速、鲁棒性高的无人机航拍图像识别算法。具体研究目标包括：通过构建航拍图像的图模型表示获得地物目标间拓扑关联知识，联合多拓扑结构并进行学习与优化，实现多任务提升算法下的分类识别模型。研究结果有望为视频监控、场景标注等应用提供理论方法和关键技术支持，使无人机航拍图像识别研究更好地在低空遥感测量和监测等方面发挥作用。

本课题以计算机视觉、机器学习相关理论和方法为基础，以无人机航拍图像为研究对象，围绕图像内容的图模型表示与匹配、拓扑结构选择与学习两个关键科学问题，重点研究航拍图像目标之间拓扑关联知识发掘，多拓扑结构学习与优化，基于多任务提升算法的分类模型构建等内容，进而自动、高效地发现航拍图像中目标之间拓扑结构等几何模式，并实现航拍图像目标的自动实时识别。.本研究首先采用YOLO v3目标检测算法及YOLACT图像分割算法构建航拍图像拓扑图模型并进行子图挖掘；然后采用格拉斯曼流形嵌入算法对拓扑图进行向量化映射，从而使拓扑图子图能够对比计算，筛选具有高判别度低冗余度拓扑结构；最后将筛选后的拓扑结构和航拍场景类别作为神经网络输入输出，搭建深度学习模型，通过对深度学习模型训练优化，完成航拍图像识别任务。.课题围绕图模型表示与匹配、拓扑结构选择与学习两个关键科学问题，以无人机为典型应用平台，研究航拍图像的实时分类识别方法。在课题推进过程中产生了图像增强去雾、无人机避障和无人机目标检测跟踪领域内重要研究成果。研究结果有望为视频监控、场景标注等应用提供理论方法和关键技术支持，使无人机航拍图像识别研究更好地在低空遥感测量和监测等方面发挥作用。