High-resolution Remote Sensing Image Building Extraction Based On U-shaped Full-convolutional Neural Network

With the continuous expansion of the scale of the city and the increasing number of buildings,the importance of the extraction of buildings to urban planning is also increasing.With the continuous development of remote sensing technology,the current extraction of buildings is mostly done based on high-resolution remote sensing images.At present,the main technologies of building extraction from high-resolution remote sensing images can be summarized as spectral and morphological index extraction methods and deep learning extraction methods.With the advancement of deep learning technology,the accuracy of high-resolution remote sensing image extraction methods based on deep learning has been greatly improved,and deep learning has gradually become the mainstream technology for building extraction based on high-resolution remote sensing images.In view of the fact that the current full convolutional neural network is more inclined to the extraction of the main part of the recognition feature,the feature extraction part of the low receptive field area is ignored,resulting in the lack of the model for the detection of the edge of the feature.The bilinear interpolation algorithm in the sampling stage cannot extract the outline of buildings in a regular manner,so that the extraction results of buildings from high-resolution remote sensing images based on deep learning require people to make secondary annotations and modifications before they can be put into application,which greatly reduces the efficiency of interpretation.The main research contents of this thesis are as follows:(1)In order to improve the extraction accuracy of the model for high-resolution remote sensing image buildings,this thesis proposes an improved CBAM＿ResNeXt＿SPP＿Unet network based on the Unet network.Using ResNeXt as the basic network,the features of reducing feature correlation through grouping convolution operations enrich the abstraction extracted by the model.Features: Introduce the SPP pyramid pooling network,enrich the main feature information of the building through multi-scale feature fusion,and extract the main features of each part of the abstract feature map of the building to achieve high-precision extraction of the backbone of the building;Fusion CBAM volume The cumulative attention module captures the main features of the edge of the building outline while enriching the main features of the main part of the building,and improves the segmentation effect of the model for the building outline;(2)In view of the problem of sawtooth and noise in the segmentation results of building outlines,the ConvCRF end-to-end model is introduced,and the building extraction results are inferred multiple times through the hidden Markov property model.The object and the building edge outline are further segmented,so as to achieve the purpose of noise reduction and building outline edge refinement;At the same time,in view of the problem that the upsampling operation of the fully convolutional neural network cannot regularly extract the outline of the building,an improved Douglas Peucker algorithm is proposed to correct the outline edge of the building extraction result such as rotation and translation.Linearization and verticalization of corners to achieve regularized extraction of buildings.The research results show that the CBAM＿ResNeXt＿SPP＿Unet fully convolutional neural network proposed in this thesis is superior to the traditional fully convolutional neural networks such as Unet,Deep Labv3 Plus,FCN,etc.It has a certain degree of improvement in the average precision and average intersection and union,and its MIOU reaches 0.8750.MPA reaches 0.9593,which is higher than the accuracy of cutting-edge and classic semantic segmentation models in the task of high-resolution remote sensing image building extraction;the introduced ConvCRF model does not need to significantly increase the interpretation time of high-resolution remote sensing image buildings,but also effectively reduces interpretation.The resulting edge jaggedness and noise,refine the building edge segmentation results,and effectively pave the way for building regularization.The CBAM＿ResNeXt＿SPP＿Unet model connecting the end-to-end ConvCRF model achieves MIOU of 0.8773 and MPA of 0.9601 on the test set,which not only improves the The model’s segmentation accuracy for buildings also improves the segmentation effect of building edges;the improved Douglas Peucker image post-processing algorithm can further effectively improve the regularization effect of buildings on the basis of deep learning interpretation results,thereby reducing manual secondary Labeling work.The research results can provide a certain theoretical reference for urban planning,land resource management,ecological environmental protection and other fields.