High-Resolution Image Building Extraction Based on CNN-Transformer Multi-Scale Features and Boundary Awareness

In the task of building extraction from high-resolution remote sensing images, challenges remain due to large variations in building scales, confusion between buildings and complex backgrounds, and oversmoothing of boundary details. Existing deep learning methods based solely on convolutional neural networks (CNNs) or Transformers often fail to simultaneously capture local geometric details and global semantic consistency. This study proposes a dual-branch building extraction network that combines CNN-based multi-scale spatial features with Transformer-based global semantic features. An attention-guided multi-scale feature fusion module is designed to alleviate feature conflicts and to suppress complex background noise. A boundary-aware loss and a multi-scale deep supervision strategy are introduced to enforce geometric constraints on building contours and intermediate features. Experiments on the Gaofen-7 (GF-7) dataset show that the proposed method achieves an IoU of 0.8065 and an F1 score of 0.8929, outperforming the compared methods. When using only 50% of the training samples, the model still attains an IoU of 0.7701, indicating good stability and application potential under limited labeled data. The method enables fine extraction of buildings from high-resolution remote sensing images and provides technical support for geographic information applications such as land-use survey, urban planning, and disaster assessment.