Remote sensing road segmentation and breakpoint repair based on deep learning

Abstract

The current high-resolution remote sensing image road extraction methods mostly use an end-to-end network model to predict road, which is helpful to the global feature expression of road extraction, but ignores local feature express. Therefore, in order to better balance the global information and local information extracted by the road, in this paper, we use SEResNext to extract rich semantic information from low to high channel to construct the encoder, and use the global mixed feature module (GMFM) to upsample and recover to the original image size to construct the decoder, in which dense skip connections promote semantic information fusion. In addition, we use polynomial fitting and morphological-based methods to repair road breakpoints. Experiments show that compared with similar networks, the accuracy and completeness of road extraction in this paper have been significantly improved, the precision achieves 88.765%, the recall is 84.224%, and the achieves 86.435%.