Automatic Building Extraction From High Resolution Remote Sensing Image Based On Multi-scale Feature Fusion

Remote sensing images have huge feature information and rich semantic information.In recent years,Chinese infrastructure has developed rapidly and the process of urbanization has gradually accelerated.It is of great significance to plan cities and draw maps through remote sensing images.At the same time,China is also faced with the current situation of rapid aging of population and annual decrease of newborn population.Therefore,how to estimate population density through building extraction is also of great value.Although previous traditional building extraction methods have achieved some results,traditional methods are mostly based on artificial design features such as shape,color,texture and so on.In most cases,they can only be applied to a specific building type,so they cannot achieve good results on a large range of data sets to meet the practical application needs,let alone the requirements of rapid intelligent automatic extraction.Thanks to the rapid improvement of computer hardware in recent years,computer computing power and computer vision has improved rapidly,and the subject of remote sensing has also developed rapidly.There has been a lot of research on building extraction in the past few years,due to the complexity of the texture of the building and the background of the image,high-precision segmentation of buildings from high-resolution remote sensing images is still a challenging task.Since building extraction is a binary classification task,the U-shaped structure of the network is more friendly to this task.Therefore,this paper improves the existing network LinkNet model and proposes a building extraction network model HRLinkNet based on multiresolution feature fusion of high resolution remote sensing images to complete the task of building extraction.Using ResNet18 as the backbone network to extract different levels of building features,it will utilize the dilated convolution pyramid to capture the spatial context information,merge the features of different levels multiple times and combine it with the features of the deepest upper sampling and upper sampling step by step.Predictions and multiple supervisions are carried out at each stage to make full use of features of different resolutions.Aiming at the problem of insufficient generalization ability of remote sensing image models in different domains,domain migration is used to transform the style of remote sensing images,exploring the possibility of remote sensing image conversion,and improving the generalization ability of models.Two models of Cycle GAN and VAE-GAN are used to perform image conversion,respectively using domain X to domain Y and domain Y to domain X dual learning,and a variational autoencoder to map the features of different domains to the shared space.Thus,the image conversion of the two data sets is realized,which improves the generalization ability of the HRLinkNet model.The experimental results show that :(1)Compared with other common building extraction methods,HRLinkNet,which combines multi-resolution features while maintaining high resolution,has better effect,can suppress noise and fine edge?(2)The building extraction effect of remote sensing image data sets in different domains is greatly improved by generating antagonism network transformation.