Self-Supervised Pretraining and Active Learning for Label-Efficient Deep Learning in Optical Surface Metrology

Abstract

Labeled training data is the primary bottleneck for deploying deep learning in precision optical surface metrology, where expert annotation is expensive, time-consuming, and subject to inter-observer variability. This study proposes a label-efficient deep learning framework for optical surface metrology that combines self-supervised pretraining on large volumes of unlabeled measurement data with active learning-based sample selection for annotation, dramatically reducing the number of required labeled samples. Built upon the deep learning methodologies established by Huang, Tang, Liu, and Huang (2026) in optical metrology and the 4D thermal imaging approach of Huang, Yang, and Zhu (2023), the framework first pretrains convolutional encoders on unlabeled optical measurement data using a contrastive learning objective, learning rich representations of surface geometry and thermal patterns without any manual labels. Subsequently, an active learning module selects the most informative unlabeled samples for manual annotation, focusing annotation effort on the data points that will most improve model performance. Evaluated across three representative optical metrology tasks—thermal image reconstruction, phase unwrapping, and surface defect detection—the framework achieves within 5% of fully supervised performance using only 8–12% of the labeled dataset. Active learning sample selection outperforms random selection by 31% in label efficiency, and self-supervised pretraining provides a 2.3× speedup in convergence during fine-tuning. The proposed framework provides a practical pathway toward data-efficient deep learning deployment in precision optical metrology with limited annotation budgets.