Building Polygon Extraction from High-Resolution Remote Sensing Imagery Using Knowledge Distillation

Building polygons plays an important role in urban management. Although leveraging deep learning techniques for building polygon extraction offers advantages, the models heavily rely on a large number of training samples to achieve good generalization performance. In scenarios with small samples, struggle effectively represent diverse structures and handle complexity introduced by background. A common approach enhance feature representation is fine-tuning pre-trained model dataset specific task. However, process tends overfit task area leading loss knowledge from dataset. To address this challenge enable inherit while characteristics paper proposes distillation-based framework called Polygon Distillation Network (BPDNet). The teacher network BPDNet trained containing samples. student was available target learn provides guidance during network, enabling it under supervision knowledge. Moreover, improve buildings against backdrop complex context, characterized fuzziness, irregularity, connectivity issues, employs Dice Loss, which focuses attention boundaries. experimental results demonstrated that addresses problem limited integrating It accurately identifies alleviates boundary fuzziness issues.