Research On Deep Learning Method For Building Segmentation Of Remote Sensing Images

Buildings are important artificial objects in remote sensing images and usually play an important role in the analysis of optical remote sensing images.Semantic segmentation of buildings in high-resolution remote sensing images is an important research work in the field of remote sensing image processing.With the continuous progress of computer,image processing and deep learning technologies,the use of these technologies to quickly and accurately segment buildings in remote sensing images has important significance and value in academic and engineering aspects.Thanks to the development of high-speed computing hardware,deep learning methods have developed rapidly in recent years as well as extensive research and applications.Convolutional Neural Network(CNN),as the most typical neural network,has been deeply researched in the field of computer vision,making image segmentation possible in deep learning.As an important method for semantic segmentation in the field of deep learning,Fully Convolutional Network(FCN)and its subsequent improvement methods are the basic methods for building segmentation in remote sensing images.At the same time,post-processing methods such as Conditional Random Field(CRF)and image filters have also played an important role in optimizing the segmentation results of remote sensing image buildings.However,due to the characteristics of remote sensing images and the requirements for the mask integrity of the segmentation results,the existing segmentation methods and post-processing methods have certain flaws.According to the characteristics of remote sensing images,this thesis proposes a deep learning remote sensing image building segmentation framework,which consists of a two-branch segmentation network that can simultaneously predict building masks and edges and a domain transform conditional random field.The domain transform conditional random field can effectively improve the performance of the building segmentation network of remote sensing images through the edge optimization mask predicted by the two-branch network.Experimental results show that the method proposed in this thesis can effectively segment the buildings in remote sensing images.This thesis mainly carried out the following research:(1)This thesis first investigates the current research status of existing remote sensing image building segmentation methods at home and abroad.The common remote sensing image segmentation methods and the application of related technologies in the field of deep learning in remote sensing image segmentation are analyzed.It also summarizes the research difficulties of deep learning methods in the field of remote sensing image segmentation.(2)In-depth analysis and research of existing deep learning image semantic segmentation technology.At the same time,basic operations in deep learning such as loss function,activation function,upsampling,etc.are introduced.After that,various post-processing methods used in the field of image semantic segmentation were studied.The main direction of the method in this thesis is clarified.(3)The core research work carried out in this thesis is to apply a Domain Transform Filter to a conditional random field to construct a domain transform conditional random field.Therefore,this thesis also reforms the building segmentation network so that it has two decoder branches to simultaneously detect the mask and edge of the buildings in the remote sensing image.The domain transform condition random field uses the edge prediction map output by the segmentation network to optimize the mask of the segmentation network.The domain transform condition random field satisfies the end-to-end requirements of general deep neural networks.Therefore,the domain transform conditional random field can be trained jointly with the segmentation networks to improve network performance.(4)The method proposed in this thesis was tested on the Wuhan University Building Dataset.This article utilizes the overall accuracy and Io U as the objective indicators of experimental evaluation.The experimental results show that the two-branch segmentation network can complete the tasks of semantic segmentation and edge extraction of buildings in remote sensing images at the same time,and the domain transform conditional random field can use building edge information to improve the building segmentation effect of remote sensing images.The experiments prove that the proposed method achieves better building segmentation effects in remote sensing images than the current representative method.