Non-Local Feature Search Network for Building and Road Segmentation of Remote Sensing Image

Building and road extraction from remote sensing images is of great significance to urban planning. At present, most building models adopt deep learning semantic segmentation method. However, the existing methods did not pay enough attention feature information between hidden layers, which led neglect category context pixels in pixel classification, resulting these two problems large-scale misjudgment buildings disconnection extraction. In order solve problem, this paper proposes a Non-Local Feature Search Network (NFSNet) that can improve accuracy roads, help achieve accurate By strengthening exploration layer information, it effectively reduce large area misclassification process segmentation. Firstly, Self-Attention Transfer (SAFT) module proposed, searches importance on channel dimension, obtain correlation channels. Secondly, Global Refinement (GFR) introduced integrate features extracted backbone network SAFT module, enhances map obtains more detailed output. The comparative experiments demonstrate proposed method outperforms state-of-the-art methods, model complexity lowest.