Building Recognition And Contour Normalization Of Remote Sensing Image Using Deep Convolution Neural Network

As one of the most important types of artificial targets in a basic geographic database,the information on the geography,shape,and color of buildings has great demands in the government administration,the civil departments,and the business sectors.These requirements are different and the information sometimes cannot be communicated with each other,so extracting building information 1s a common task for remote sensing applications and geographic information systems.In addition,as various types of houses increase,decrease,and change with urban planning and construction,such information needs to be updated in time.However,manual surveys to perform automatic updates of buildings in basic GIS databases is time-consuming and labor-intensive.The amount of data used to extract buildings from human means is limited and collection quality is inconsistent.On the other hand,with the practical application of high-resolution remote sensing data,the use of remote sensing technology for rapid,large-scale building extraction gradually shows its advantages.Among them,military detection,urban planning,statistical survey,disaster emergency assessment and other fields need to quickly and accurately extract artificial targets such as buildings from remote sensing images,and to carry out relevant applications.Therefore,the automatic extraction of building from remote sensing images gradually developed into a mainstream method for urban modeling,acquiring and updating information in the basic geographic information system database.Although some advances have been made in the detection and extraction of buildings,these algorithms still suffer from too many manual interventions,poor robust,difficulties in building multi-constraint models,and high computational complexity.Problems such as these are difficult to meet the needs of practieal applications.Therefore,for massive image data,how to quickly and automatically achieve accurate extraction of building outlines is of great significance.In recent years,deep learning theories and methods have developed rapidly.In the field of image processing,convolutional neural networks have achieved surprising effects in image recognition that have been difficult to achieve with conventional classification algorithms.Compared with the difficulty of mathematical expression and implementation caused by the human defined image features,this method can learn abstract,essential,and advanced features from a small amount of preprocessing or even raw data,and it has a certain degree of invariance for translation,rotation,scaling,or other forms of deformation.It has been widely used in the fields of license plate detection,face detection,text recognition,target tracking,machine learning,and computer vision.In this paper,we propose an strategy for the extraction of large-scale building from images,based on the convolutional neural network method,eombined with prior knowledge.It includes a series of steps such as rough extraction,building shape detection,shape matching,etc.and finally acquire the precise outlines of the buildings.It opens up a new idea for the automatic extraction of buildings.The study mainly contains:(1)Building rough segmentation.Image segmentation was taken to segment the image into several regions of interest(ROIs).A regression model was constructed using supervised learning technology,Which was used to eompute the likelihood to be a building for a ROI,and to give the decision by a proper threshold.(2)Building contour refining.A priori shape library was established,and a classification model was constructed to recognize the building shape contained in the roughly extracted ROI.Then area matching method of active contour model based on the prior shape const:raint and the edge matching method based on the set of contour points were used to retrieve the model parameter to the identified shapes of buildings to get the fine outlines.Conclusions are as follows:(1)Using the convolutional neural network method to detect building and recognize the shape in remote sensing images has a high accuracy.When the building image is recognized,the method that use multiple rotated outer rectangles to obtain ROI images for a ROI and select the maximum from it can effectively overcome the larger deviation of building orientation in training samples.(2)The area matching method of active contour model based on the prior shape constraint and the edge matching method based on the set of contour points are used to match the model parameters of the prior shape and the building contour in the image,and the building contour can be extracted accurately,even in the case of a fixed occlusion.