Few-shot concrete surface extraction for automated UAV-based building inspection using vision foundation models

Although unmanned aerial vehicle (UAV)-based building inspection requires automated extraction of concrete regions for end-to-end structural damage detection, existing automated inspection workflows lack this preprocessing step. This paper presents a few-shot learning pipeline that reformulates pixel-wise segmentation as a mask-level binary classification task, enabling site-adaptive concrete-region extraction with minimal annotation and no site-specific retraining. The pipeline pairs a frozen segmentation foundation model for mask generation with a frozen self-supervised vision transformer for mask-pooled features, training only a lightweight residual multilayer perceptron classifier. On two UAV datasets, it matches or exceeds supervised baselines trained under the same annotation budget, reaching 86.3% mean intersection over union on a complex multi-material dataset with ten labeled images. In real UAV crack inspection, restricting detection to the extracted concrete regions removes false alarms on non-concrete structures, increasing the crack F1 score from 3.5% to 27.9% and demonstrating direct field benefit for automated inspection.