A Study Of Building Extraction Method In Complex Scene Based On Deep Learning

With the rapid development of urban construction process,the surface coverage changes with each passing day.Buildings become the most easily changed part of geographic data.As an important component of geographic information database,they play an important role in urban planning and construction,urban expansion analysis and disaster early warning evaluation.Based on high-resolution remote sensing image is the main way to obtain building data.There are many building extraction methods,and semantic segmentation method based on deep learning is the best one.In high-resolution aerial image data,buildings are more regular,clear boundaries,and building extraction is relatively simple.When using semantic segmentation method to extract buildings,by improving the ability of the encoder to obtain image features,Residual U-Net network can improve the accuracy of building extraction.However,the scene of high-resolution satellite remote sensing image is more complex,high-rise buildings are inclined,and the boundary of small low and dense buildings is fuzzy,which is not conducive to extraction.Only by acquiring more image features can't effectively improve the accuracy of building extraction,By adding the expansion convolution module and fusing the deep and shallow features of the image,the context information,global information and local feature information can be obtained,so as to improve the accuracy of building extraction results.In this paper,high-resolution aerial image WHU data set and high-resolution satellite remote sensing image ZHUHAI data set are made for experiments.The results show that:(1)Residual U-Net obtains rich image feature information through the encoder network of deeper convolution neural network,which can effectively improve the extraction accuracy of buildings.95.40% of the building extraction accuracy are obtained through WHU data set experiment,So as to improve the accuracy of image segmentation;(2)Building extraction in complex scenes of high-resolution remote sensing images can't effectively improve the accuracy only by adding image feature information.By fusing shallow feature information and using dilated convolution module,local information can be obtained on the basis of global information,and more perfect context information can be obtained,which can effectively improve the accuracy of building extraction in complex scenes,By building an fusion of multi-feature improved PSPNet model,the accuracy of building extraction in complex scenes of high-resolution remote sensing images can be effectively improved,which makes the extraction results of high-rise buildings and small low density buildings complete and clear,and the accuracy of building extraction in complex scenes can reach 80.35%;(3)The fusion of multi-feature improved PSPNet model is suitable for the extraction of urban villages in complex scenes of high-resolution remote sensing images,and its extraction accuracy is 82.67%.This model is suitable for the extraction of urban villages in complex scenes of high-resolution remote sensing images,and has the value of popularization and application.