Building Extraction Method From High Resolution Remote Sensing Image Based On CNN

The fast and efficient extraction of ground object information is the key to the application of massive remote sensing data,and the extraction of buildings in highresolution remote sensing images is one of the research hotspots.Although the building detection and extraction technology has made some progress at present,most of them still require human participation.The related theories and methods are not mature enough to meet actual engineering needs.In recent years,deep learning methods have made great progress in the field of image recognition and processing.According to the requirements of extraction of general building contour and specific building information,this paper studies different types of convolutional neural network models for the purpose of image semantic segmentation and target detection,builds training data set,improves the structure of convolutional neural network,establishes extraction models of general buildings and specific buildings,and improves the performance of the high-resolution remote sensing image building extraction efficiency and accuracy.The main work of the paper is as follows:(1)Research on convolutional neural network method.The basic principle and structure of the convolutional neural network are studied,the characteristics of the local receptive field,weight sharing,and multi-core convolution of the convolutional neural network are analyzed.The principle of the transfer learning method is studied,and the basic structure of the Tensor Flow framework is analyzed.The existing neural network model for building extraction from remote sensing images is studied,and its existing problems are analyzed.(2)Research on building extraction of high-scoring remote sensing image based on Deep Lab v3+.Aiming at the problem that the building boundary segmentation accuracy in the existing remote sensing image building extraction model is not high,a remote sensing image building extraction method based on the improved Deep Lab v3+ network is proposed.The thesis improves the basic model of Deep Lab v3 + network,uses the cubic convolution method instead of the original bilinear interpolation method for upsampling processing,and introduces a transfer learning mechanism to accelerate the speed and accuracy of model training.This paper uses the Shanghai dataset in Space Net data for network training.The experimental results show that the improved model can segment the buildings and background in the image well,and the average intersection ratio obtained is 78%.The extraction accuracy of buildings in remote sensing images is effectively improved.(3)Research on target detection and information extraction of oil storage tank.In this paper,a typical circular building-oil storage tank is taken as the research object,and the rapid extraction of specific types of buildings and the acquisition of accurate parameters are studied.Using the GF-2 image as the data,a remote sensing image sample set of the oil storage tank target was established,and the Faster R-CNN model was optimized by using the strategies of improving the feature extraction network and optimizing the regional suggestion network,and the oil storage tank target extraction model is established.The research shows that the m AP value is increased by 6.39%compared with the original Faster R-CNN,and the high-precision detection of the oil storage tank target in the remote sensing image is realized.Based on the sub-pixel analysis of detecting the shadow of the tank body,the length of the shadow can be accurately extracted,and the height of the tank can be calculated through the shadow space geometric relationship of the remote sensing imaging.In combination with the tank radius obtained by the Hough transform,a rapid estimation of the volume of the tank is achieved.Experiments show that the method proposed in the paper can quickly detect the target of the oil tank,extract the shadow of the oil tank and estimate the height of the oil tank.The average relative error of the estimated volume of the oil tank is2.37%,indicating that the method is highly feasible.