Road Extraction From High Resolution Remote Sensing Image Based On Lightweight Convolutional Attention Network

The road is the main part of modern transportation,and it is very important to manage and update the road information in the GIS database.Remote sensing image data has rapidly become the main data source for automatic road network extraction.It can provide high-precision ground information and more complex background details.It can also carry out large-scale road monitoring,which brings a lot of room for improvement to road extraction.However,at present,visual interpretation is still the main way to update the road,which is costly and requires a lot of time The time and resources of automatic road extraction have an impact on the task.The road in remote sensing image has a large intraclass difference and a small inter-class difference of the special-shaped structure area,that is,the complex road structure features,and the road will be visually blocked by building shadows and trees,resulting in the difference between the two features is not obvious,and then can not accurately reflect the road information.Although traditional road extraction methods can obtain high-quality road extraction results in some regions,they often rely too much on the selected features.These problems increase the difficulty of automatic road extraction from remote sensing images.The rapid development of deep learning has opened up a new research direction for remote sensing image processing.It can extract different levels of image features through learning in the samples,so as to learn more about the research objects in the input image.In this paper,the deep learning method is used to extract roads from high-resolution remote sensing images automatically,quickly,and efficiently:(1)In view of the discontinuity of road extraction caused by the occlusion of trees and buildings in remote sensing images,the attention mechanism is introduced into a convolutional neural network to expand the receptive field of the network.Combined with the context information of the road,features that contribute a lot to the task are selected to suppress irrelevant semantic ambiguity,which can significantly improve the performance of the remote sensing image road extraction model.(2)In order to further improve the road extraction of remote sensing images and lighten the network model,this paper proposes a lightweight convolutional attention neural network model.The model adds spatial channel attention mechanism to a convolutional neural network,weights features from two dimensions of space and channel,selects road features that are useful for tasks,suppresses or weakens irrelevant semantic features,and reduces the over-segmentation of low-level features by the network,so as to improve the performance of road extraction model.In addition,in order to make it possible to apply the remote sensing image road extraction model in real life,the depth separable convolution lightweight model is introduced into the network,so that it can predict the road in the remote sensing image faster without affecting the actual performance.(3)Programming a "high-resolution remote sensing image road automatic extraction system",design visual interface,convenient for users to operate.The system integrates all the remote sensing image road models involved in this paper.The system can intuitively compare the road extraction effects of various algorithms,and achieve fast and accurate road extraction from high-resolution remote sensing images.