Adaptive Local Prototypes for Cross-Domain Few-Shot Defect Segmentation in Intelligent Robotic Inspection

Abstract

Most industrial defect segmentation methods rely on extensive pixel-level annotations and are limited to known defect categories and domains. To address this limitation, this paper investigates cross-domain few-shot defect segmentation, where meta-knowledge learned from a data-rich source domain is transferred to unseen target domains using only a few annotated support samples. We propose an Adaptive Local Prototypes Network (ALPNet). It first formulates support-query matching as an entropy-regularized optimal transport problem to generate robust multi-scale correlation maps. Highly correlated support features are then used to construct task-adaptive local prototypes, which further produce a prior mask for query segmentation. The correlation maps and prior mask are jointly decoded without target-domain training or fine-tuning. Experiments on five target datasets and a practical steel-rust inspection system demonstrate strong cross-domain generalization. ALPNet achieves $\mathbf{6 1. 5 7} \%$ and 69.78% mIoU on ISIC 2018, and 44.81% and 54.43% on unseen steel-rust segmentation under one-shot and five-shot settings, respectively.