Building Extraction From High Resolution Remote Sensing Images Based On Deep Learning And Line Detection

In recent years,the global Earth observation technology has developed rapidly,and the Earth observation system has become more and more perfect,as represented by the domestic high-score series images and the Geo-Eye satellite images of the United States.High-resolution remote sensing images and their derivatives have been widely used in environmental protection,agricultural investigation,disaster prevention and mitigation,urbanization research,marine development and other fields.With the rapid rise of artificial intelligence,deep learning,cloud computing and other emerging technologies,how to efficiently and accurately extract useful information from high-score remote sensing images has become the key whether high-score images can be widely used.As the main place of human production and life,building area is a kind of target which plays an important role in both military and civil fields.Therefore,the extraction of high-resolution image building objects has important research value.Based on depth learning and line detection,this paper presents a high accuracy method for building object extraction in high-score images.The main work of this paper is as follows:(1)this paper downloads high resolution remote sensing image data containing large number of building targets from Google Earth,and establishes a standard target detection data set called WHURS_Building including building detection training data set and test data set.(2)Based on the region-based full convolution neural network(R-FCN)and fully convolutional residual network(ResNet-50),building target feature learning and training are carried out,and the building detection template parameters are obtained.Building detection is carried out in the detection test data set to obtain the preliminary extraction region.(3)According to the characteristics of building object,this paper presents a method of accurate extraction of building area based on line detection.In the preliminary extraction area of the deep learning building detection,the information of the line segment is obtained by LSD,and the lines are filtered according to the length of the line segment to remove the background straight line noise.Then,according to polar coordinates,close line clustering and merging are carried out to eliminate redundant line information.Finally,the effective line information is converted into binary data,and the coordinate data set of the building area is obtained,and the minimum outer rectangle of the line segment in the local region is obtained by using the minimum external rectangle algorithm of the point set in the two-dimensional coordinate system.That is the exactly extracted area of the building.By using the high resolution remote sensing images building test data set established in this paper,the results show that the method in this paper is superior to other algorithms in the accuracy and recall of building detection in high-resolution images and the precision of extraction of building area is further improved.