Research And Implementation Of Fast Road Extraction From Satellite Imagery

Road extraction is one of the significant tasks in remote sensing field.In recent years,with the development of deep learning,semantic segmentation has been applied to the task of road extraction.Road extraction based on semantic segmentation mainly uses satellite image and its corresponding mask to train the segmentation model,and then extracts the road contour and center line by post-processing.However,the main problem of the current road extraction system is that its inference speed is limited,which means it cannot deal with the massive satellite image data in limited time.Therefore,this paper tries to solve the efficiency problem by improving efficiency on semantic segmentation algorithm and engineering techniques.From the point of view of deep learning algorithm improvement,one possible solution is to introduce real-time semantic segmentation.Compared with traditional semantic segmentation,real-time semantic segmentation not only emphasizes absolute accuracy,but also focuses on efficiency.Based on the introduction of real-time semantic segmentation,this paper also studies the methods to improve the accuracy and generalization performance of real-time semantic segmentation model,including model distillation.Besides,for road extraction task,we have proposed a novel vectorization algorithm based on square measurement for line geometry objects.From the perspective of engineering improvement,this paper proposes a pipeline of road extraction system,which uses parallel operation to reduce the time consumption of bottleneck steps in the road extraction process.Meanwhile,the half precision inference deployment framework is used to improve the inference speed of semantic segmentation model.In summary,this paper proposes an efficient road extraction system with a lightweight road extraction algorithm based on semantic segmentation and an overall framework to solve the inference speed problem on road extraction.Improvements are made on both algorithm and engineering.Experiments are made to validate our proposed methods.