G2LDIE: Global-to-Local Dynamic Information Enhancement Framework for Weakly Supervised Building Extraction From Remote Sensing Images
Jiaxing Sun, Wei He, Hongyan Zhang
Abstract
Image-level weakly supervised semantic segmentation (WSSS) methods have gained prominence in remote sensing image building extraction tasks, primarily due to their cost-effectiveness in manual annotation. However, owing to the intricate details present in remote sensing building images, the pseudolabels generated from existing image-level weakly supervised methods often encounter issues of incorrect activation and unclear boundaries. In this article, we propose a global-to-local dynamic information enhancement (G2LDIE) framework. This framework effectively extracts global information from remote sensing building images and supplements local details through a local information enhancement (LIE) module, generating more accurate pseudolabels. Additionally, we propose a dynamic label guide strategy (DLGS) to enhance model consistency in category representation across various scale images. To address the challenge of unclear building boundaries issue in pseudolabels, we introduce a segment anything model (SAM) postprocessing (SPP) method, which can better correct the boundaries of building images at different resolutions while reducing computational costs. Extensive and detailed experiments on three datasets confirm that our framework can generate refined pseudolabels and outperform other image-level weakly supervised methods in terms of accuracy and generalization performance in building extraction.