Building Extraction And Contour Optimization Methods In Remote Sensing Images

Massive high-resolution remote sensing images provide data supports for acquiring building information,which is a critical component of urban infrastructure.By extracting and analyzing buildings from remote sensing images,the spatial distribution and shape information of buildings can be captured,which has important research and application significances in optimizing urban planning,promoting digital cities,assessing disasters and reconnoitring military targets.However,due to the interference of irrelevant objects around buildings,existing building extraction methods often suffer from pixel omission and mislabeling,leading to some defects such as edge serration and irregular shapes in the extracted buildings.To address this issue,this paper focuses on building extraction and contour optimization using high-resolution remote sensing images as the research object,and proposes a remote sensing image building extraction algorithm and two building contour optimization methods.The main contributions are listed as follows:1)A deep learning-based method for building extraction in remote sensing images is proposed,which utilizes multi-scale and spatial attention mechanisms to explore the multi-scale spatial contextual information of remote sensing images and achieve effective segmentation of buildings.Specifically,this study proposes a building segmentation model based on multi-scale spatial attention(MSA-Res M),which consists of three main parts.To begin with,a Residual Encoder(Res-Encoder)is employed to semantically encode the remote sensing image and prevent network degradation and gradient vanishing issues through residual connections.Subsequently,a Multi-scale Spatial Attention Module(MSAM)is proposed,which integrates spatial attention with multi-scale mechanisms to capture spatial contextual positional relationships at different levels and mine richer semantic representations,thereby optimizing the processing of buildings of different sizes.Finally,the semantic features extracted by ResEncoder and the multi-scale spatial features extracted by MSAM,are fed into Res-Decoder to decode features for pixel-wise building extraction results.Experimental results demonstrate that the proposed model effectively extracts the outlines of buildings in remote sensing images,and optimizes existing deep learning segmentation algorithms.In addition,we conduct ablation experiments to further confirm the effectiveness of the Multi-scale Spatial Attention Module introduced in this study.2)An optimization algorithm based on corner recognition and localization is devised to extract building profile.By extracting corners to excavate the prominent geometric structural characteristics of buildings,the proposed method can efficiently represent building shapes with a small amount of data.Our method introduces the orthogonal complete function system Vsystem to fit building profile and utilizes cluster analysis to recognize and localize corners by analyzing the fitting results.In particular,the optimization method consists of three parts.First,the MSA-Res M model is employed to extract building regions from remote sensing images.Then we deploy a contour tracking algorithm to obtain contour point sets of buildings.Second,the V-system is used to fit building contours.Third,the cluster algorithm is utilized to analyze the skewness and relative spatial position of the fitted line segment sets to achieve corner recognition and localization.Finally,the corners are connected with line segments to regularize the building boundaries.Experimental results show that the proposed method can not only preserve important structural features of buildings,but also effectively restore the straight edge features of building outlines,thus achieving significant improvements of building contour boundaries.3)An optimization algorithm based on V-system fitting is proposed for building outline extraction,which is based on the fact that buildings edges often are straight,and that curve fitting can intuitively represent the data distribution,making it naturally advantageous for representing the overall profile of buildings.By using linear V-system that adapts to straight edges to fit buildings,the proposed method can obtain building boundaries with smooth edges.Specifically,our method is composed of three main parts.First,the cluster analysis and linear V-system are used to extract building outline corners.Second,we optimize the parameters of the whole discrete point sets,map the corner parameters to V-system nodes,as well as transform the non-corner parameters to the corner parameter space to ensure that V-system can accurately reconstruct the corner features of buildings.Finally,based on different features of corner and non-corner points in building shape representations,we construct the continuity constraint condition for V-system basic function nodes,and further solve it using the Lagrange multiplier method to obtain the final fitting curves.Experimental results demonstrate that the proposed method can effectively represent the overall shape of the buildings and preserve detailed characteristics of corners,which is able to reveal real features of buildings and represent complex surface features.