Building Extraction In Airborne Oblique Remote Sensing Image Based On Instance Segmentation

Oblique remote sensing images contain ground object information from multiple views.Buildings are one of the main types of ground objects in remote sensing images.Instance-level extraction can obtain information such as the location and geometry of each building,which is convenient for detailed information analysis of individual buildings.The instance segmentation method can obtain a large number of instance-level building areas through pixel-level annotation information.The research results of building extraction from airborne oblique remote sensing images based on instance segmentation method can be used for the maintenance and statistics of buildings,and provide important support for the update of national geographic database information.The instance-level extraction of buildings depends on the large-scale data with detailed annotation information.In addition,the characteristics of the oblique remote sensing image itself also bring challenges to the extraction of buildings.In summary,there are three main problems: 1.The lack of large-scale annotated datasets that can be used for building extraction research in multi-view oblique remote sensing images.2.The complex background information,occlusion and adjacent buildings in oblique remote sensing images affect the accuracy of building instance-level extraction.3.The pixel-level annotation for the target is a huge work.In view of the above problems,this paper studies from three aspects:1.Aiming at the lack of datasets for related research,this paper constructs two airborne oblique remote sensing image data.The MSB-Dortmund dataset and MVB-Zurich dataset.The MSB-Dortmund dataset contains buildings with different size,and the MVB-Zurich dataset contains buildings from multiple views.It provides data support for further analysis of multi-size building extraction in oblique remote sensing images and building extraction from different views.2.In the case of supervised learning,aiming at the complex background information,occlusion and adjacent situation in the oblique remote sensing image affect the accuracy of building instance-level extraction,this paper proposes an instance segmentation framework FRNet.Aiming at the background problem,this method uses the feature enhancement method based on the self-attention mechanism to suppress the background information and enhance the characteristics of the building itself.Aiming at the problem of occlusion and adjacent situation,the two-layer segmentation branch based on the constraint of remote sensing boundary loss function is used to predict the output of the building.The effectiveness of the method is proved by experiments on the MSB-Dortmund and MVB-Zurich datasets.3.In the case of weakly supervised learning,aiming at the problem that the pixel-level annotation for the target is a huge work,this paper proposes a weakly supervised instance segmentation framework GMBox based on bounding box supervision.This framework extracts the instance-level building through the bounding box of the building.Firstly,the multi-scale gradient prior fusion is used to guide the multi-instance learning task.Then,the mask correction module is used to improve the mask predicted under weakly supervised conditions and improve the accuracy of weakly supervised building monomer extraction.The effectiveness of the method is proved by experiments on the MSB-Dortmund and MVB-Zurich datasets.