Road Extraction From High Resolution Remote Sensing Images Based On Dual Spatial Attention Mechanism

In the high-resolution remote sensing image,the road shows the complexity of the background,the vulnerability to occlusion,the long-span natural connectivity and the variability of the road grade,which makes the existing deep learning methods limited by the lack of long-distance dependence acquisition ability of the encoder-decoder structure,as well as the loss of spatial resolution caused by the encoder structure,resulting in the problem of not extracting of road results in blocked and narrow sections.To solve this problem,this paper attempts to use a dual spatial attention network for road extraction from the perspectives of capturing effective long-distance dependence and reducing the loss of spatial information.The specific work is as follows:(1)This paper uses the improved U-Net as the basic network,uses Res Net as the feature extractor in the encoder part of the basic network,and changes its original structure by using the concept of small convolution kernel instead of large convolution kernel,so as to reduce the number of down sampling and reduce the loss of spatial information;In the decoder part,an asymmetric path with the encoder part is constructed to reduce the flow of low-level rough noise information into the final classification layer and alleviate the interference of noise information on the recovery process of road structure information.(2)A dual spatial attention mechanism with global attention module and local refinement module is constructed,in which the Disentangled Non-local module is used as the global attention module to capture the global context dependence,and the local refinement module is designed with the help of the concept of multi-scale convolution,so that the module can further refine and aggregate the remote context information.Integrate them into the appropriate location in the infrastructure network structure,so that it can capture the road context information with global and local consistency,and help the network establish the topological relationship between broken sections.The experimental results on Massachusetts road dataset and LSRV dataset show that this method has significant advantages in maintaining the integrity and accuracy of road extraction.In addition,in order to prove the strong generalization ability of this method,large remote sensing images are used for model transfer evaluation.Compared with other methods,this method obtains the best value on the two effective indicators of F1-score and IoU.