Search papers, labs, and topics across Lattice.
This paper introduces rough path signature-guided geometry augmentation (RPS-GA), a novel method for enhancing few-shot industrial surface defect detection by leveraging the second-order signature responses of Canny edge contours. By aggregating these responses into a spatial map through two fusion operators, SIG-AUG and SGAA, the approach significantly improves detection performance in scenarios with limited labeled data. Evaluations on NEU-DET and PCB-Defect datasets demonstrate that RPS-GA can substantially raise mean average precision (mAP) scores, particularly in few-shot settings, outperforming baseline methods without requiring modifications to the YOLOv8n detector.
RPS-GA boosts few-shot defect detection performance by transforming edge contours into informative spatial maps, achieving up to a 70% increase in mAP with minimal labeled data.
Few-shot industrial defect detection remains difficult for standard supervised detectors, which achieve poor performance on boundary-dominated industrial defects. This paper proposes rough path signature-guided geometry augmentation (RPS-GA), a geometry-aware approach in which Canny edge contours are treated as ordered planar paths whose truncated second-order signature responses, especially the antisymmetric L茅vy-area term, are aggregated into a spatial map that highlights boundary-related structure through two fusion operators, SIG-AUG and SGAA. The approach is evaluated on NEU-DET and PCB-Defect under a few-shot protocol with 5, 10, 20, or 50 labeled images per class, using an unmodified YOLOv8n detector throughout. Compared with the baseline, RPS-GA delivers large gains when supervision is limited, although the margin shrinks as more labels become available. On NEU-DET, SIG-AUG raises 10-shot mAP@0.5 from 0.341 to 0.583, whereas on PCB-Defect, SGAA improves 10-shot mAP@0.5 from 0.086 to 0.299 and yields usable detection at 5-shot where the baseline fails entirely. These trends are confirmed by multi-seed evaluation across independent random partitions. Overall, the results indicate that second-order path-signature geometry offers a practical way to strengthen few-shot industrial defect detection without meta-learning or detector redesign.
