Swin-ResUNet+: An edge enhancement module for road extraction from remote sensing images

Road extraction from remote sensing images is very important in navigation, urban planning, traffic management and other fields. Deep learning methods have achieved great success in computer vision tasks. Therefore, road extraction from remote sensing images using deep learning methods can significantly improve the road extraction accuracy. However, these methods generally have problems such as low road extraction accuracy, slow training speed, high computational complexity, and poor road topology connectivity. In order to solve the above issues, we propose a Swin-ResUNet+ structure and use the new paradigm Swin-Transformer to extract roads in remote sensing images. Specifically, we construct an Edge Enhancement module based on residual connection and add this module to each stage of the encoder, which can obtain the edge information in remote sensing images. Based on the Edge Enhancement module, we propose a Swin-ResUNet+ structure in order to better capture the topology of roads. On the Massachusetts road dataset, our model has the least computational cost with only less than one percent accuracy decrease. On the DeepGlobe2018 road dataset, our model not only has the least computational complexity but also achieves the highest values of mIOU, mDC, mPA and F1-score. In a word, Swin-ResUNet+ obtains a much better trade-off between accuracy and efficiency than previous CNN-based and Transformer-based methods.