Edge-Aware Dual-Stream Dynamic Fusion Network for Building Extraction in Remote Sensing Images

The extraction of buildings from high-resolution remote sensing images faces significant challenges due to complex urban scenes, including spectral ambiguities, geometric diversity, and boundary uncertainties. Despite progress in deep learning-based methods, buildings' varied shapes and sizes complicate edge extraction, making it difficult for existing approaches to balance the global context with local details and leverage edge priors effectively, especially for adjacent buildings. In response to these issues, we propose an edge-aware dual-stream dynamic fusion network called EA-DSFNet. It follows the encoder-decoder paradigm, where the encoding component utilizes a dual-path setup integrating convolutional neural networks (CNNs) and Transformer modules. First, we design a feature reuse and fusion (FRF) module that dynamically integrates local information extracted by the CNN with global contextual features extracted by the Transformer at the encoder stage to achieve a comprehensive multi-scale feature representation. Second, a feature selection gate (FSG) is developed to adaptively emphasize discriminative features during cross-branch fusion while suppressing redundant information. Aiming to improve the model's utilization of building edge information, the decoder network introduces a multitask edge-directed decoder (MTEDD). This architecture optimizes edge extraction and segmentation tasks jointly through hierarchical supervision, enabling the decoder to fully assimilate edge information. Additionally, the foreground and background enhancement module (FBEM) explicitly utilizes geometric priors to guide segmentation, thereby improving the boundary accuracy of the segmentation results. We perform extensive testing across three standard reference datasets to assess the performance of our method. The experimental results demonstrate that EA-DSFNet surpasses current leading techniques in segmentation.