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

Road information is an important part of planning or repairing road traffic,and it is also the basis for road traffic to gradually develop toward intelligence and automation.With the development of remote sensing technology,remote sensing images have gradually become the main means of obtaining road information.How to quickly,efficiently and automatically obtain road information from remote sensing images has become a hot topic in the field of remote sensing.In recent years,deep learning algorithms have made great progress in combining remote sensing big data with computer image processing technology to obtain road information,but there are still certain shortcomings.From the analysis of the existing algorithms,it can be seen that there are still some phenomena such as loss of detailed features,blurred edges,discontinuities of road sections,and errors in topological structure in the information obtained by deep learning algorithms.Therefore,based on the existing algorithms,this article proposes two methods of improvement: one is based on the improved U-Net algorithm of dilated convolution to extract road information from remote sensing images;the other is based on the improved U-Net algorithm of dense connection extract road information from remote sensing images.The experimental results obtained by training the model under different data sets show that the accuracy indicators of Recall,Precision,and F1-score of the dilated convolution U-Net and the densely connected U-Net model are as high as 82% or more,even higher in the Cas Net Data Set data set.Achieve more than 92%.The specific content and achievements are as follows:(1)Established a remote sensing image road data set.In order to enhance the generalization and universality of the network,this article self-made the remote sensing image road data set of Nanming District and Baiyun District of Guiyang City,introduced its process and data enhancement.At the same time,collect public remote sensing image road data sets for comparison and verification.Experiments show that the generalization and universality of the network depend on the diversity of the sample data set.The more types of sample data,the stronger the adaptability to the same type of data,and the better the extraction effect.(2)Improved U-Net algorithm based on two methods of dilated convolution and dense connection.1)Using U-Net as the basic network,extract the middle layer convolutional network and change it to a multi-scale feature extraction module constructed by dilated convolution with different expansion rates to obtain multi-scale information of road features.The experimental results show that it improves the network's acquisition of road detailed features and the improvement of edge fuzzy features.2)Replace the convolutional network of the U-Net network with the dense connection method of the dense connection network to strengthen the propagation between features and enhance the utilization rate of features.The experimental results show that the network can improve the discontinuities of road sections to a certain extent,thereby enhancing the topology of the network.(3)The prototype system for automatically extracting roads from remote sensing images.In order to facilitate the use of users in actual production,the models trained by the above two improved algorithms are integrated and packaged in a user interface,which includes file opening of selected images,road prediction and display of results.