부산대학교 도서관

Extracting buildings from high-resolution remote sensing imagery (HRSI) is of great significance to emergency management, land resource utilization, and analysis, as well as city planning and construction. However, due to the complex backgrounds and diverse appearances and different sizes of buildings in HRSI, most existing methods for automatic building extraction are difficult to obtain strong building feature representation from low-level and high-level features. Furthermore, existing research mainly focused on regional accuracy, whereas less attention was paid to the description of building boundaries. In this article, MSB-Net, an end-to-end neural network, is proposed to address these issues. A multiscale feature fusion module (MSFFM) is designed to capture and fuse multiscale features. A local branch (LB) constructed by the MSFFM and position attention, is used to obtain long range of context information between different positions and extract the essential features of buildings (e.g., shapes, edges) from low-level features. And a global branch (GB) is designed to use the MSFFM and channel attention to enhance high-level features. Therefore, our method can not only obtain information on building-related attribute categories, but also capture the rich context information in channel dimensions. The boundary enhancement and completion module take the output of the GB and LB as input to search for the missing parts and details of buildings to improve the segmentation accuracy and boundary quality. Our method is tested on two public building datasets and achieves superior classification performance.