Research On Road Extraction Based On Remote Sensing Images

Automatic Road Extraction from remote sensing images remains an important computer vision task.With the rapid development of deep learning in the direction of image segmentation,the performance of road extraction tasks based on high-resolution remote sensing images has been significantly improved.However,the existing Road Extraction methods and Road Centerline Extraction methods have certain shortcomings.On the one hand,segmentation-based road extraction methods usually rely on intermediate non-graph representations(semantic masks)and a post-processing heuristics module.The pixel-wise classification approach also results in a lack of model awareness on road integrity and topology so that the neural network often predict road network with noise and low accuracy.On the other hand,the road centerline extraction task based on graph methods predict the structure of the road network graph.The generation process of road network is often iterative,which leads the neural network model to pay more attention to local information and lack of consideration of global information.Therefore,this paper focuses on the tasks of Road Extraction and Road Centerline Extraction based on remote sensing satellite images in open scenarios.First,this paper extracts pre-training schemes for roads in different geographical areas based on road vector data,so that the model has better generalization ability and implicitly improve the topology prediction ability of the model in the fine-tuning stage.Second,this paper proves the effectiveness of context modeling in the road extraction task and further improves the global reasoning ability of the model based on the improved SwinTransformer structure for the Road Extraction task.And prove the effectiveness of the model on the DeepGlobe public dataset.Thirdly,for the open and complex remote sensing image scene,this paper proposes a heuristic post-processing method to improve the smoothness of the prediction mask.Finally,this paper innovatively regards road network map prediction as a discrete sequence prediction problem,and uses Transformer Decoder as a sequence generation model.At the same time,the multi-task learning is used for extract the road centerline as an auxiliary task,so that the model has high segmentation performance and improves the topological connectivity of the predicted road.