Building Extraction From Remote Sensing Images Based On Context Parsing

Remote sensing technology is a kind of detection technology which uses various sensors to detect electromagnetic wave information radiated and reflected by long-distance target.Building extraction from remote sensing image is one of the important applications of remote sensing technology.Building is the carrier of human life and production activities.It is of great significance to study the distribution of buildings for the study of population aggregation and dispersion.In addition,building extraction has important application in urban planning,land investigation,military reconnaissance and so on.In the past,the building extraction of remote sensing images was mainly completed by classifying the features of artificial design.However,due to the influence of spectral images and the diversity of buildings distribution,the efficiency of these methods is low,and affected by the factors such as sunlight and shelter,and the generalization performance of feature extraction is poor,which can not meet the needs of real scene.With the development of deep learning in recent years,more and more people try to apply it to the building extraction of remote sensing images.This paper starts from the context semantics of building remote sensing images,and uses semantic segmentation algorithm to analyze the features.The specific work is as follows:(1)The traditional SegNet has many parameters,and it is easy to appear the problem of gradient disappearance and feature extraction ability weakening in the process of building remote sensing image training.This paper proposes a Separable Residual SegNet(SR-SegNet).SR-SegNet introduces an improved residual block in the down sampling stage,which alleviates the problem of gradient disappearance in the process of model training.At the same time,it can extract deep semantic information,and introduce deep separable convolution to reduce model parameters and training time.The experimental results show that the parameters of SR-SegNet are reduced by 2/3 and the segmentation accuracy is improved by 3.84%.(2)The overall texture and background of high-resolution remote sensing image of buildings may be very complex,such as spatial resolution,shape and scale of buildings.In the field of deep learning,in order to extract more detailed building features from remote sensing images,more complex convolution operations and larger network models are usually used to segment buildings.On the one hand,the segmentation accuracy has been improved obviously,but the segmentation speed is not satisfactory.In this paper,the improved feature pyramid pooling and encoder-decoder are combined to propose Feature Residual Analysis Network(FRA-Net)considering the segmentation accuracy and speed.The experimental results show that the mean intersection over union of evaluation index is 1.87% higher than that of the classical Fully Convolutional Network,and the prediction speed is 3.57 times.