BuildNext-Net: A Network Based on Self-Attention and Equipped With an Efficient Decoder for Extracting Buildings From High-Resolution Remote Sensing Images

As is known, the accurate extraction of buildings from high-resolution remote sensing images has become a pivotal objective. In some complex scenes (e.g., there will be objects with a similar spectral texture to buildings in the image, trees and shadows will obscure the buildings, etc.), the existing models cannot accurately recognize the buildings. To address this series of challenges, we propose a new method, BuildNext-Net, which consists of TransNext-EMAM blocks, upsampling convolution modules, context feature enhancement blocks (CFEBs), and multiscale depthwise convolution blocks (MSDWCBs). The encoder consisting of TransNext-EMAM blocks is used for feature extraction and outputs the generated feature maps of each layer to CFEB through skip connections. In the feature reconstruction stage, CFEB can receive the jump-connected feature maps and the feature maps obtained from upsampling, which improves the network’s capacity to comprehend and localize the target objects and image details. MSDWCB can further enhance the multiscale feature extraction capability to achieve the effect of suppressing irrelevant regions to capture multiscale salient features. It effectively solves the challenge of combining local and global information in complex scenes. It also enhances the robustness of the network in extracting buildings in complex scenes. Our method has been extensively experimented on the WHU building dataset, the Massachusetts building dataset, and the Inria building dataset. The intersection over union metrics on these three datasets are 91.21%, 76.12%, and 81.42, improving 1.06%, 1.55%, and 2.60%, respectively, compared with other state-of-the-art methods.