| The extraction of buildings from high-resolution remote sensing images is currently a relatively avant-garde research direction,especially the rapid development of artificial intelligence technology and deep learning models have continuously promoted its vigorous development in recent years.It provides more solutions for the general survey and planning of urban and rural buildings,the evaluation of rooftop solar resources,geographic information systems update and buildings change detection of remote sensing images.Firstly,we introduce the related knowledge of building extraction from remote sensing images and then some basic theories and technologies used were introduced in this paper.In addition,we also introduce the used dataset and its construction process.Secondly,the features of the five adopted classic segmentation models are analyzed and introduced in this paper.The five used classic segmentation models are FCNs,U-net,Deeplabv3+,YOLACT and Mask R-CNN respectively.Thirdly,we also summarize the basic segmentation mechanisms,advantages and disadvantages of each model.Remote sensing images with the labeled data are selected for verification and pixel-level building extraction from building images are realized.Combined with the segmentation results from different building images,the performances of five classical segmentation algorithms are analyzed from three evaluation dimensions.Which are the execution time,memory occupation and the accuracy of our segmentation algorithm effectively.At last,we summarize the advantages of some segmentation algorithms and specific application scenarios which provide good reference for selecting the appropriate segmentation algorithm combined with specific application scenarios.In the process of applying Deep Convolution Neural Network to process highresolution images building extraction,due to these improvements of deep networks,the interference of images background information,the unreasonable and scientific multiscale semantic information fusion which lead to problems of slow training and fitting speed,inaccurate edge building segmentation and there are some holes in large-scale segmented buildings.In order to solve these above problems,characterize the differentiated remote sensing features of different cities and the problem of imbalanced data in this model,we employ the deep learning framework of Pytorch based on the constructed remote sensing images processing dataset,and adopt the Vortext Pooling module and Dual Attention Mechanism Module to improve together multi-scale feature fusion ability of Deeplabv3+.Moreover,we adopt the mixed parameter management method of binary and floating-point numbers,combine with the multi-task learning algorithm to realize the extraction goal of buildings based on high resolution remote sensing images.It also solves the problems of inaccurate segmentation of edge buildings and holes among large-scale buildings in semantic segmentation,speeds up the fitting and convergence speed of the algorithm model.What’s more,the accuracy and effect of building extraction are improved effectively.Compared with other comparison models,various evaluation metrics and segmentation effect are better than other similar models,which has important reference value for building segmentation and extraction based on remote sensing images.Finally,we predict the research technology and development trend of building extraction from remote sensing images in the future. |
