Direction-Emphasizing Transformer for Road Extraction From Optical Remote Sensing Imagery

This work contributes an U-Shaped encoder-decoder network model for highly accurate road extraction from optical remote sensing imagery. By designing a direction-emphasizing transformer based on the combination of strip and traditional square convolutions in the decoder, the specific characteristics of road objects such as long-range connectivity, and geometric topology are robustly synthesized. The encoder is combined with a large selective kernel-based encoder that can strongly encode multilevel features, the relationships of global-local information in the image, the challenges of the road extraction task are effectively solved. Comprehensive experiments on benchmark datasets showed that the proposed network model outperforms previous state-of-the-art methods in terms of IoU and F1-score indexes.