Road Extraction From High-resolution Remote Sensing Images Based On Semantic Segmentation

High-resolution remote sensing road image is a road image based on high-quality remote sensing technology.With the rapid development of satellite technology,extracting roads from remote sensing roads has extremely important economic value and scientific research significance in automatic map updating,urban road planning,intelligent vehicle navigation,etc.How to accurately and quickly separate roads from remote sensing images has become a difficult problem in front of people.In this thesis,various difficulties in the process of road extraction from remote sensing images are studied.The main work is as follows:(1)To solve the problem that road extraction is difficult due to land cover,building cover,tree occlusion and so on in the process of road extraction from high-resolution remote sensing images,this thesis designs a multi-task key constraint module,which constrains the connection weight ratio between roads by calculating the semantic constraint angle between road pixels,so that the connection between road pixels is more transcendental,which improves the difficulty of road extraction due to the difficulty of road identification in images and improves the road extraction ability of the model in complex scenes.(2)In the remote sensing image,the road target is long and narrow,which usually presents a certain degree of continuity,which leads to the change of road background after a certain area,which makes the background information of the road complicated and produces noise of different sizes at the edge of the image scene change.Therefore,this thesis adds a dual attention mechanism module,which makes the context information of the remote sensing road image focus on extracting the information of the road itself in two different dimensions: location and channel,while filtering other redundant information to a certain extent.(3)In the process of down sampling,the road feature information will be lost with the decreasing resolution of feature map,which leads to inaccurate road extraction results.In this thesis,a multibranch cascaded cavity space pyramid module is constructed,which can extract features of different scales from multiple branches while keeping the resolution of feature map unchanged,thus improving the accuracy of road extraction.Combining the above research contents,this thesis proposes a new road target extraction model based on semantic segmentation for high-resolution remote sensing images.A large number of experiments have been done on public dataset Massachusetts and private dataset RSR.The experimental results show that our proposed model has higher accuracy and good robustness in road extraction.