基于全卷积神经网络的无人机影像建筑物提取

Fully convolutional network-based building extraction of image from unmanned aerial vehicle

针对无人机影像建筑物自动提取时精度低、边缘精细化程度不足导致水利移民征地过程中建筑物面积统计不准确的问题,为提高无人机影像建筑物自动提取精度,利用基于跳层连接的全卷积神经网络进行无人机影像建筑物自动提取,结合无人机数据生产过程中产生的DSM对建筑物初步提取结果进行后处理,得到更加精细化建筑物的边缘。结果显示,利用基于跳层连接的全卷积神经网络进行无人机影像建筑物提取,平均Kappa系数为0.90,平均查准率为0.93;当前常用的分类模型Deeplab网络,Kappa系数为0.67,查准率0.70;Segnet网络的Kappa系数为0.75,查准率为0.79。相较于Deeplab网络Kappa系数提高了0.23,查准率提高了0.23;相较于Segnet网络,Kappa系数提高了0.15,查准率提高了0.14。利用DSM对初步提取结果进行后处理,处理后Kappa系数为0.92,查准率为0.94。对比初步提取结果,Kappa系数提高了0.02,查准率提高了0.01。结果表明,结合DSM与全卷积神经网络的无人机影像建筑物提取方法具有更优越的提取精度。

Aiming at the problem of the inaccurate statistics of the building area during the land expropriation for the construction of water conservancy project caused by the low accuracy and insufficient edge refinement of the building image automatically extracted by unmanned aerial vehicle,the automatic extraction of the building image from UAV(unmanned aerial vehicle)is made by means of the skip-layer connection-based fully convolutional network(SLC-FCN),so as to improve the automatic extraction accuracy of the building image from UAV,and then the post-process is carried out on the preliminary extracting result of building in combination with the DSM(Digital Surface Model)data generated during the data generation of UAV,thus the more refined edge of building is obtained.The result shows that the mean Kappa coefficient and the mean precision ratio are 0.90 and 0.93 respectively for extracting building from UAV image through the SLC-FCN.At present,the Kappa coefficient of the frequently-used Deeplab network is 0.67 with the precision ratio of 0.70,while the Kappa coefficient of Segnet network is 0.75with precision ratio of 0.79.Compared with Deeplab network and Segnet network,the Kappa coefficients are increased by 0.23 and 0.15 and the precision ratios are increased by 0.23 and 0.14 respectively for extracting building from UAV image through the SLC-FCN.The processed Kappa coefficient after the post-process of the preliminary extracting result with DSM is 0.92 with the precision ratio of 0.94.Compared with the preliminary extracting result,the Kappa coefficient is increased by 0.02 with the increase of the precision ratio of 0.01.The study result indicated that the method combining DSM with the fully convolutional network for extracting building from UAV image has more preferable extraction accuracy.