Research On High Resolution Remote Sensing Image Building Segmentation Based On Deep Learning

In the emergence of urban disasters,rapid reconstruction of city maps is particularly important.The rapid development of global positioning system and satellite positioning technology provides this possibility for the rapid reconstruction of urban buildings.Urban high-resolution building images can be obtained from nano-satellite remote sensing satellites and unmanned aerial vehicles,and conventional high-altitude satellites.Images have various advantages such as high resolution,fine identification of objects,and strong timeliness.At present,traditional methods such as threshold,region,and edge segmentation have harsh segmentation conditions and certain limitations.In recent years,deep learning and other methods used for semantic and instance segmentation have mostly targeted low-and medium-resolution pixel-by-pixel segmentation.The method cannot make full use of the characteristics of high-resolution and rich underlying semantic information of remote sensing images,and it difficult to obtain more accurate boundary segmentation results.In order to response the above problems,this paper studies the segmentation of buildings under high-resolution remote sensing images,so as to accurately extract the building Mask segmentation and contour polygon vertex information.First of all,by analyzing the advantages and disadvantages of the pyramid feature fusion network and the path fusion network in the field of semantic segmentation,a method based on the multi-scale layer-by-layer attention mechanism feature fusion network remote sensing image building semantic segmentation based on ResNet is proposed.Multiple features of different scales generated by ResNet are sent to the GFNet module layer by layer for fusion.Finally,feature maps with diversity and the same size of the input data are used to vote for buildings and non-buildings in remote sensing images.Finally,on the third-party data set Inria Aerial Imagery for Labeling Dataset GFNet is used to evaluate the remote sensing image building segmentation.GFNet has a data set resolution of 5000*5000,which improves the performance of the PANet and UNet segmentation network by about 1.2 compared with the path fusion network.%.In terms of the performance accuracy index of the verification set,GFNet is second only to the top IctNet with a performance of 0.2%.Compared with other top 90 networks,it has an improvement of 0.3%-12%,ranking second.Secondly,in view of the problem that the method of image segmentation based on deep learning is a pixel-by-pixel raster map,and the three-dimensional reconstruction of GIS data of urban buildings requires vertex,edge,polygon and other data,this paper proposes a feature based on DenseNet A semi-automatic building segmentation method for remote sensing images based on the detection of the vertices of the pyramid building contour polygon.This method encodes the center of gravity of the building,launches several rays at the center of gravity,predicts the position of the polygon vertex in a regression method,and returns to the vertex position of the building polygon.In this paper,the number of experimental rays is 36 and 72 for experimental comparison.Comparing the same network,the results on the training visualization graph of Dice and Recall show that the increase in the number of rays improves the accuracy of polygon vertex detection and segmentation to a certain extent.In summary,through the study of convolutional neural network path fusion network and feature pyramid network in the field of image semantic segmentation,this paper solves to a certain extent the problem of boundary accuracy segmentation in remote sensing image building segmentation tasks.The segmentation is not accurate enough and the speed is slow,which not directly generate the polygon vertex information of the city's 3D reconstruction vectorized data.This paper proposes two methods for semantic segmentation of buildings in remote sensing images.The two methods have distinct characteristics.Different segmentation methods be selected according to different requirements for building segmentation.The layer-by-layer fusion method based on ResNet directly segment buildings in the form of masks,and has a high segmentation accuracy,which is suitable for quickly and directly obtaining the scene of direct segmentation of buildings.Based on the DenseNet feature pyramid remote sensing image building contour polygon vertex segmentation method,it can directly generate the polygon vertex information of the threedimensional reconstruction of urban buildings,or directly obtain the polygon vertex information of the building to perform three-dimensional reconstruction and mark the location of certain key points of the building.