Building Extraction Based On Convolutional Neural Network

With the fast development of remote sensing,satellites provide a large amount of remote sensing data.Buildings are one of the most important features in high resolution remote sensing images,and it is an important issue to extract buildings from these complex and massive data automatically and efficiently.Solving this problem helps us with urban planning,and so on.At present,some researches have been made on the detection,recognition and extraction of buildings in high resolution remote sensing images in China and foreign countries.However,the spatial structure and texture characteristics of buildings are very complicated,and the environment in which the buildings are located is complex.Buildings may also be blocked by trees.These characteristics have caused certain difficulties in extracting buildings,and cannot fully meet actual needs.The purpose of this research is to extract buildings from high resolution remote sensing images,which use the public data set as experimental data.The specific research content includes the following points:Firstly,this thesis reviews the current research of building extraction and convolutional nerual networks in China and foreign countries.We analyze and summarize the current situation of building extraction,introduce the basic knowledge of convolutional nerual networks,and explain the advantages of building extraction using convolutional nerual networks.Secondly,traditional algorithms cannot effectively extract buildings.To solve this problem,this thesis proposes a semantic segmentation method based on convolutional nerual networks.This method is based on UNet model,and added feature fusion branches to improve the poor segmentation of buildings.An image post-processing module is added to solve the problem of non-smooth edges of segmentation results to obtain better building extraction.Thirdly,for the problem of only classifying each pixel in semantic segmentation,this thesis proposes an instance segmentation method based on convolutional nerual networks.This method identifies each building as an independent individual.This method is based on MS RCNN model,and added atrous convolution to increase the receptive field.At the same time,a fusion path is added to FPN network to make full use of features,and the Soft NMS algorithm is used in RPN network for better region proposal.Fourthly,in order to solve the problem of poor edge extraction in building extraction for convolutional neural network models,a fully connected conditional random field is added after networks to obtain better results.Finally,these methods were tested on the same data set,and compared with traditional methods to verify the effectiveness of methods in this thesis.The research showed that the methods in this thesis can extract buildingsautomatically,efficiently and accurately.Both the improved UNet and the improved MS RCNN can achieve better results than the original model.This research has certain reference value for building extraction.