Automatic Road Extraction Of High-Resolution Remote Sensing Image Based On Deep Neural Network

As a popular research method in recent years,deep learning is extending its application to the interdisciplinary field.High-resolution remote sensing image carries a wealth of ground object information.As an important task in the field of computer vision recognition,road automatic extraction of high-resolution remote sensing image is of great research value in many practical scenarios such as vehicle navigation,intelligent transportation,city planning and map drawing,and has been concerned by experts and scholars at home and abroad.Compared with traditional road extraction methods,the advantages of deep learning-based automatic road extraction mainly lie in that it can use the learning ability of deep network to identify and extract road feature information more intelligently.In this paper,an end-to-end high-resolution remote sensing image road segmentation network is proposed,which can automatically learn multi-level and multi-scale road feature information.To solve the problem of road segmentation in high-resolution remote sensing images,the A-DLinknet network model proposed in this paper introduces a channel-spatial dual attention module on the basis of D-Linknet segmentation network.This module sets a new weight distribution layer through mask,marks out the road features in the image,and the network obtains the marked road area through learning and training,thus forming attention.It can effectively enhance the characteristic information of the road and solve the problems of"false check","missed check"and"false check"in the original segmentation network to some extent.In addition,a super parameter weight loss function based on dice+bce was constructed.According to the prior weight ratio of 1:1,2:1,3:1,4:1 and 5:1,F₁-score was used as the evaluation index of the segmentation network.The performance of the segmentation network reached the best state when the super parameter weight ratio was 4:1 through several experiments.Five segmentation networks,FCN-8s,U-Net,Deep Labv3+,DANet and D-Linknet,were applied to the road extraction task of high-resolution remote sensing image,and their training process was analyzed by several evaluation indexes.Due to the influence of image quality,shooting time and other factors,the performance of remote sensing image segmentation network on real data sets may decrease significantly.Using a semi-supervised generation antagonism model,A-DLinkNet is taken as the generator network of the model,and the full convolutional network as the discriminator network.The feedback mechanism is generated to minimize the set multi-task loss and space cross entropy loss.Under the condition of cross data set,this model can significantly improve the road extraction effect.Under the same data set condition,the road extraction performance of the generated network is improved to some extent.