Research On Road Extraction From Remote Sensing Image Based On Fully Convolutional Neural Network

The extraction and precise segmentation of remote sensing image roads is conducive to the accurate monitoring and path analysis of the country's roads across the country,and is also conducive to the construction and maintenance of roads in remote areas.High-resolution remote sensing images have high resolution,complex terrain and are susceptible to vehicle shadows,which increase the difficulty of accurate road extraction.The research finds that the existing remote sensing image road extraction algorithm has the problems of artificial participation and low extraction accuracy.Therefore,it is very important to study how to extract roads from high-resolution remote sensing images quickly,accurately and efficiently.This paper absorbs and draws on the research results of previous scholars to improve the accuracy of remote sensing image road extraction.Based on the FCN model,two simple and effective road extraction algorithms for remote sensing images are proposed:(1)For the semi-automatic remote sensing image,the road extraction algorithm requires manual intervention and the road extraction accuracy is not high.In this case,a multi-source input feature fusion full-convergence neural network remote sensing image road extraction algorithm is proposed.Firstly,the Prewitt operator is used to perform edge detection and gray processing on the data set to form a new data set.Then the edge detection,grayscale processing and color map are generated as multiple input sources to generate three parallel convolutions.In the network,the three parallel convolution results are feature-fused,and after deconvolution,the label map is classified at the pixel level to obtain a road segmentation model with multiple input sources.The results show that compared with the traditional remote sensing image road extraction algorithm,the algorithm of this paper can automatically and efficiently extract the road from remote sensing images.(2)Loss of feature information and resolution reduction caused by layer-by-layer convolution and downsampling of the traditional FCN model,which leads to the gradual reduction of features learned by the network,resulting in an unsatisfactory road extraction effect.In this paper,a remote sensing image road extraction algorithm is proposed based on multi-volume neural network for porous convolution and feature fusion.First,the porous convolution is used to replace the standard convolution of the last two stages in the original full convolutional neural network,which not only increases the receptive field,but also retains more characteristic information;then,the results of the convolution of each stage are feature-fused.Both the global feature information and the local specific feature information are added.Without increasing the number of parameters,the receptive field of the convolution kernel is enlarged,so that the model can learn more feature information and improve the segmentation accuracy without increasing the calculation amount.The experimental results show that the model achieves better segmentation accuracy in high-resolution remote sensing image road extraction.