Building Extraction From Remote Sensing Image Based On Convolutional Message Passing Network

Building extraction based on high-resolution remote sensing images is of great strategic significance to social economy,urban construction and disaster resistance.How to get the edge information of buildings from high resolution images(high resolution images)more accurately and efficiently is a hot topic in the current research.Due to the various structures of buildings and the occlusion or interference of the key features of buildings in high-resolution images,the traditional algorithms for building extraction from high-resolution images generally have the disadvantages of low extraction accuracy and long extraction time.Therefore,how to extract buildings from high-resolution images with high accuracy is an important topic.With the progress of hardware devices,the field of deep learning has developed rapidly.Deep learning algorithm has the advantages of high accuracy and strong robustness in image processing,so deep learning can be introduced in high-resolution image building extraction to achieve higher extraction accuracy.Therefore,this paper analyzes the research status of building extraction at home and abroad.Aiming at the current situation that the existing methods only extract the building edge More than that the building internal geometric structure,this paper proposes a Multi-level coding decoding network building extraction method based on message passing network.This paper includes the following two aspects:1)A method of building extraction based on convolutional message passing network is proposed.The whole algorithm is mainly composed of feature extraction module and edge verification module.The feature extraction module uses the residual network Res Net50 as the backbone network to generate candidate regions;the deep residual network with extended convolution is used to extract the feature map for subsequent coding operation;the edge verification module uses the decoder to up sample the output feature map,and finally uses a full connection layer to output the confidence score at the end of the decoder.Finally,the binary cross entropy loss function is added to increase the robustness of the model;2)Based on Space Net data set and WHU data set In this paper,PPGNet,L-CNN and UNet++ algorithms are used to compare with the results of this paper.Through the analysis of experiments,the results show that the accuracy,recall and F1 score of region extraction are superior,which indicates that the Multi-level coding decoding network based on message passing network has some advantages.