Learning discriminative universal background knowledge for few-shot point cloud semantic segmentation of architectural cultural heritage

Few-shot point cloud semantic segmentation (FS-PCSS) is crucial for architectural heritage digitization, as it segments unseen categories with minimal annotations. However, existing methods focus on target features while neglecting background modeling, compromising segmentation performance through artifacts and diminished feature clarity in non-target regions. To address these issues, we propose LUBK-Net, a novel framework that strengthens background feature alignment to mitigate the feature deviation caused by the domain gap. Specifically, a Background Attention Module (BAM) learns a universal background prototype using an adaptive background loss, while removing irrelevant features to suppress background noise. In addition, a Prototype Contrastive Learning (PCL) further enhances foreground-background discriminability. Experiments on the Self-built Ancient Architecture and the publicly accessible Architectural Cultural Heritage datasets demonstrate that our method outperforms state-of-the-art approaches, offering new insights for heritage conservation.