Research On Remote Sensing Road Extraction Method Based On Improved LinkNet

Road,as the most basic feature element in remote sensing images,has always been a hot and difficult point for experts and scholars from all over the world.The diversification of the scenes in which it is located has reduced the versatility of traditional methods based on shallow features.With the application of deep learning in the field of computer vision,remote sensing image road extraction tasks have entered the era of deep learning based on deep abstract features,and based on the road itself The characteristics of continuous improvement.Among them,the D-Link Net network model that achieved the highest score in the 2018 Deep Globe Road Extraction Challenge（Deep Globe Road Extraction Challenge）is based on the original Link Net network model to optimize the road extraction results by introducing a hollow convolution module.However,the road extraction results still have problems of connectivity and accuracy.This paper studies the above two problems separately,and proposes an improved Link Net remote sensing road extraction model DD-Link Net（Densely D-Link Net）.The improvement mainly includes the following two aspects:（1）For the connectivity problem of road extraction,this article believes that the receptive field formed by the introduced cavity convolution module is not dense enough,and the feature scale obtained is too small to make full use of the extracted feature pixels.Therefore,the hole convolution module in the D-Link Net network model is improved,and the hole convolution layer connected in series in the original hole convolution module is modified into a dense connection,that is,the feature data output by each layer of hole convolution is used as After that,the input feature data of all the hollow convolutional layers allows each neuron that receives the feature data to obtain more scales of semantic information for processing,and finally the formed multi-scale features with denser coverage are fused to improve Connectivity of road extraction.（2）For the accuracy of road extraction,this article believes that the joint loss function used in the training process is caused by the inaccurate evaluation of the error.Therefore,the joint loss function of BCE loss and Dice loss in the D-Link Net network model is improved,and a new joint loss function of Focal loss and Dice loss is obtained to optimize the accuracy of the error calculation during the training process.The Focal loss forces the model to change.Learn difficult-to-classify pixels well to improve the accuracy of road extraction.Finally,on the public Massachusetts road data set,Link Net,D-Link Net,DD-Link Net proposed in this article and other semantic segmentation networks U-Net,Seg Net,U-Net++and Deep Lab V3+ are compared and tested,and extracted from the road As a result,the network model proposed in this paper is indeed better than the original network in terms of the connectivity and accuracy of road extraction.Among them,Io U is improved by 0.0365 relative to the Link Net network,0.032 relative to the D-Link Net network,and F1 is improved by 3.9relative to the Link Net network.%,which is 3.54% higher than D-Link Net.