基于DeepLab V3+深度学习的无人机影像建筑物变化检测研究

Building Change Detection Based on Deep Learning with the Application of DeepLab V3+ for UAV Image

针对城市建筑物变化检测的高时间和高空间分辨率需求,研究了基于深度学习的无人机影像建筑物变化检测方法。以大疆精灵4无人机正射影像为数据源,利用变化矢量分析与人工判别生成变化标签数据,并通过数据增强生成数据集。基于DeepLab V3+深度网络,利用生成的数据集进行无人机影像建筑物变化迁移学习,最终实现建筑物变化检测。实验结果表明,深度学习可有效用于无人机影像的建筑物变化检测,对本实验所用数据总精度达到97%以上,可为大范围城市建筑物动态检测、违章检测、损害检测等提供有力支撑。

In view of the requirement of highly temporal and spatial resolution for urban building change detection, a building change detection method based on deep learning is studied. Taking the orthophoto image of Dajiang Phantom 4 UAV as the data source, the change label data is generated by combining change vector analysis and artificial discrimination. The final data set is generated through data enhancement. Based on the Deeplab V3 + network, the building change detection is realized by taking the change magnitude image as input data for the transfer learning of building change detection. The experimental results show that deep learning can be effectively used in building change detection of UAV images, and the total accuracy of the data used in this experiment can reach more than 97%. Deep learning methods can provide a support for large-scale dynamic detection of urban buildings, illegal buildings, and building damage detection, etc.