A New Method for Road Element Extraction Based on Fully Convolutional Network

Abstract

Aiming at the problem of extracting road elements from high-resolution satellite photos, a new method based on deep learning techniques is proposed and implemented in this paper. Unlike the traditional road extraction algorithms, the new method regards road elements as semantic objects and classifies them using a well trained fully convolution network, which is constructed by adding deconvolution layers to a VGG16 network model. The experimental result indicated that the trained network model is not only able to extract roads from satellite photos, but also segment major road elements at a high accuracy rate. Especially, the method has shown better robustness in dealing with the road edges which are partially blocked by vehicles or trees than that of traditional road extraction algorithms.