A weakly supervised road extraction approach via deep convolutional nets based image segmentation

Extracting road information from remote sensing images plays an import role for many practical areas. In this paper, an approach for road extraction is proposed, in order to obtain standard road region with high accuracy. By utilizing the road design and construction specifications made by the transportation industry, the road objects are assigned into different classes. Then the corresponding task is considered as an image segmentation approach, and deep convolutional network is applied to perform pixel-level estimation to predict the ownership probability of different classes. Besides, a modification processing approach is presented to exploit the segmentation result and obtain formal road network by connecting the missing or unsmooth road subsections. Experiments on remote sensing images are performed, and show that the method is efficient for acquiring multi-type roads from complex situations.