Research On Automatic Building Extraction Method Based On Semantic Segmentation Of High-resolution Remote Sensing Image

With the rapid development of remote sensing imaging and computer technology,extracting building information from high-resolution remote sensing images containing rich spatial information has become an important reference index for evaluating urban development.However,due to the high redundancy of high-resolution remote sensing image information and the complexity and diversity of ground features,the design of automatic building extraction algorithm has become a very challenging and widely concerned research.Deep learning shows excellent performance in building extraction.However,due to the irregular shape and great difference of buildings,the existing deep learning methods are difficult to extract large-scale distributed buildings.At the same time,due to the complex occlusion,shadow and other problems in high-altitude imaging,it is also very easy to lose the details of small buildings when extracting.In order to solve the above problems,this paper fully analyzes the characteristics of Massachusetts remote sensing image and WHU satellite image data,and proposes a new depth learning extraction method from the two aspects of multi-scale information and pixel context information,which further improves the accuracy of building extraction.The main research contents are as follows:(1)A building extraction method based on multi-scale fusion of high-resolution remote sensing images is proposed.This method realizes the full fusion of multi-scale information of remote sensing images.Firstly,the feature extraction part obtains different scale features through the parallel fusion module.Then,the fusion part combines gating algorithm,improved dense feature pyramid and porous space pyramid pooling to realize multi-scale fine fusion.Finally,the feature reorganization part enhances the expression of channel key features through deconvolution and channel attention.Sufficient experiments show that this method enhances the multi-scale information fusion ability of high-resolution remote sensing images and improves the accuracy of building extraction.(2)A building extraction method based on pixel context of high-resolution remote sensing image is proposed.This method realizes the full acquisition of pixel context information of remote sensing image.Firstly,the feature extraction part obtains the pixel context features through the residual network module and multi attention module.Then,in the fusion part,the spatial nonlinear transformation element is introduced to improve the ability of feature nonlinear expression and realize feature fusion through residual connection.Finally,the feature reconstruction part enhances the acquisition of pixel context information through deconvolution and OCR decoder branch.Sufficient experiments verify that this method enhances the ability to obtain the context information of high-resolution remote sensing images and further improves the accuracy of building extraction.