Research On The Application Of Semantic Segmentation Technology Of Remote Sensing Building Images Based On Deep Learning

With the breakthrough and development of remote sensing satellite technology worldwide,the information contained in the remote sensing images is increasingly rich and clear.The texture,geometry,color and other information recorded by the images can provide indispensable underlying data for major projects such as urban planning,earthquake rescue and traffic planning.In recent years,although the research of building extraction from remote sensing image has developed,there is no perfect method to adapt to the characteristics of building morphology which changes with time,region,aesthetic and other factors,so the extraction of building from remote sensing image is still facing many challenges.Traditional methods for building information extraction from remote sensing images are usually limited by high labor cost and low accuracy.Therefore,it is of great value to apply deep learning technology to building information extraction from remote sensing images.In this paper,a Multi-scale fusion of Deformation Residual Pyramid Network（MDRP-Net）was proposed by referring to relevant literature at home and abroad,combining the characteristics of buildings in remote sensing images and the principle of convolutional neural network.Firstly,the model of Unet is improved by introducing the deep coding network and adding the path of feature transmission in the network layer.The original single serial network structure is changed into a dual path feature extraction network,which is dominated by the deep coding network and supplemented by the subsampling bypass network.Then,a pyramid network structure is introduced at the end of the coding network,which takes the deformation convolution feature as the total input and uses the residual modules to connect the different scale features to enhance the adaptability of the network to the diversity of the building contour and the variable scale.Finally,a joint up-sampling decoder is used to complete the task of decoding and recognizing the building features of remote sensing images.The results show that the two new networks can improve the segmentation accuracy based on the original Unet model,and the segmentation accuracy is the highest when the two networks are used together,which are 0.870 and 0.784 in the F₁ and MIoU evaluation indexes,respectively.On the self-made data set,the performance comparison experiment of MDRP-Net model before and after using random erasure preprocessing is compared.First,the Local Space Viewer platform was used to intercept the remote sensing images that publicly disclosed by Google,and a binary data set of buildings in the university town of Panyu District,Guangzhou was made by ourselves.Then the self-made data set was preprocessed by modifying the random erasure generation strategy.Finally,a comparative experiment was designed and three types of test areas were selected to predict the model,and a comprehensive analysis and evaluation of the prediction results were carried out.