Deep Learning For Duilding Recognition

Buildings are important grounds of urbanization.Buildings are recognized accurately,rapidly and automatically,which can provide accurate building information for the field of urban dynamic monitoring,and also have important significance for the development of urbanization planning.With the development of deep learning,image semantic segmentation has become an important research direction in the field of computer vision.The image semantic segmentation algorithm based on deep learning provides a new research idea for building recognition.The buildings in the image are the research objects of this paper.The image semantic segmentation algorithm of deep learning is a research method to classify each pixel in the image and identify the buildings in the image.This paper proposes two deep learning network semantic segmentation network models of EU-Net and DG-Net,which use these two semantic segmentation models to recognize buildings in the image.The specific work is as follows:(1)The U-Net image semantic segmentation algorithm is studied.It is found that the U-Net fixed receptive field is difficult to deal with the multi-scale of objects in the image.Based on U-Net image semantic segmentation network model,this paper adds a multi-scale pooling structure to improve the network model,and proposes EU-Net image semantic segmentation to obtain the features of buildings of different scales in the image.Through experiments,the results of EU-Net semantic segmentation model prediction are compared with U-Net.EUNet improves the recognition accuracy of buildings and determines the advantages and characteristics of EU-Net for building recognition.(2)In order to further improve the recognition accuracy of the image semantic segmentation network model for buildings,DG-Net image semantic segmentation is designed.Aiming at the space loss caused by the reduced resolution of the pooling layer,based on the characteristics of dilated convolution with increasing the receptive field and maintaining the resolution of the feature map,a dilated convolution structure with various expansion rates is proposed.The dilated convolution structure with multiple expansion rates alleviates the loss of spatial information caused by the resolution reduction caused by the pooling layer,and also improves the ability of the features of different scale buildings in the network learning image.The dilated convolution structure with multiple expansion rates alleviates the loss of spatial information caused by the resolution reduction caused by the pooling layer,and also improves the feature learning ability of the network to different scale buildings in the image.The global convolution structure is also added to the network structure,which improves the robustness of the semantic segmentation network model.In addition,the residual network structure allows the network model to fuse more contextual information.By comparing the results of DG-Net,U-Net and EU-Net experiments,DG-Net is better than U-Net and EU-Net for building recognition,which proves the correctness and feasibility of DG-Net.