Road network extraction and vectorization of remote sensing images based on deep learning

The road in the remote sensing image has the characteristics of slender and tortuous shape, complex connectivity, large road span, strong connectivity, complex ground information of the remote sensing image, occlusion, and different scales, etc. Based on these characteristics, this paper will do some change in Unet, included three parts: encoding, center feature extraction, decoding. The encoding part is aimed at problems that Unet can't extract rich feature for few samples. We proposes to use a pre-trained VGG model as an encoding module to improve network performance, reduce the risk of overfitting, and extract rich feature; the center feature extraction part, for the problems of occlusion, scale change, connectivity, etc., extracts the dense Dilation convolution module to maintain the resolution while expanding the receptive field to extract more distinguishing features; the decoding part In view of the different importance of features at different scales, a channel attention module is proposed to realize the fusion of high and low features. Through experimental comparison and analysis, the model has a significant effect on solving the road extraction problem of remote sensing images. Then the skeleton is extracted from the segmentation results and processed by denoising and straight line connection. The road network map is extracted from the skeleton map and vectorized to realize the automatic grid. Grid data vectorization.