Research On Building Extraction Method Based On Lightweight Network And High Resolution Remote Sensing Image

Building information plays an important role in urban planning,population assessment,disaster prevention and control and other applications.With the rapid development of aerospace technology and remote sensing imaging technology,the quality and quantity of remote sensing images are constantly improving,providing rich data sources for accurate acquisition of building information.The improvement of remote sensing image resolution makes the building information in the image clearer,but at the same time,along with the increasing interference of non-building details,traditional building extraction methods appear poor extraction effect,slow extraction speed and other problems.With the development of deep learning technology,building extraction methods based on deep learning technology also develop rapidly and achieve good results.However,there are still some problems in building extraction based on deep learning technology.First of all,the existing deep learn-based building extraction methods use continuous convolution operation to enlarge receptive field to obtain global information,which is easy to lose the details of building edge and lead to inaccurate building extraction.Secondly,the complex network structure contains a lot of weight parameters and computation,which limits the application of the algorithm in practical engineering.To solve the above problems,this paper proposes a lightweight building extraction method based on multi-scale feature fusion and a lightweight building extraction based on global context information.The main work of this paper is as follows:(1)A lightweight building extraction network based on multi-scale feature fusion is proposed.In this network,lightweight feature extraction network Reg Net is used as an encoder,and an improved depth-separable atrous convolution pyramid pool module is introduced behind the encoder to fully integrate multi-scale features of buildings.In the decoder structure,in order to highlight the characteristic information of the building,the channel attention mechanism is used to assign the weight of the channel and strengthen the expression of the building features before the fusion of shallow features and deep features.Finally,a lightweight decoder basic unit is proposed to further refine the segmentation results and improve the accuracy of the network.A number of comparative experiments show that this method can obtain high precision building segmentation results with a small amount of parameters and computation,and has good robustness.(2)In order to obtain more precise building contour,a lightweight building extraction method based on global context information is proposed.This method takes encoder-decoder structure as the basic framework.Firstly,the network made of Mix-Transformer modules is used as the backbone network for feature extraction,so as to fully obtain the global context information.Secondly,in order to enhance the multi-scale characteristics of the network,the feature maps of different levels in the encoder are fused with the multi-level feature fusion module proposed in this paper to realize the fusion of deep semantic features and shallow geometric features.Finally,the fused feature map is passed through the improved decoder basic unit to further enhance the extraction of building features.In this paper,sufficient experiments have verified that this method performs better in remote sensing image building extraction tasks and further improves the accuracy of building extraction.(3)The public experimental data selected in this paper are WHU satellite image data and WHU aerial image data,whose resolutions are 0.45 m and 0.3m,respectively.In order to test the building extraction effect of the proposed model in higher resolution images,this paper takes part of Haizhu District of Guangzhou as the research object and collects high-resolution UAV remote sensing images in the research area.By preprocessing images,marking building labels manually and clipping,a high-resolution building data set with a resolution of 0.1m was made by ourselves.In this paper,experiments are carried out on building data sets with various resolutions.The experimental results show that compared with U-Net,Bi Se Net V2 and Segformer networks,the two models proposed in this paper have better robustness and wider adaptability in building extraction.