Fringe property-guided deep learning method for interferogram segmentation

Accurate segmentation of interferogram is critical for optical measurements and metrology. Deep learning methods for interferogram segmentation represent a developing trend but face significant challenges due to the scarcity of annotated real interferograms. Although simulation techniques can provide alternative training data, substantial domain gaps between synthetic and real interferograms severely degrade neural network performance in practical scenarios. To bridge this gap, we proposed a fringe property-guided deep learning method that incorporates fringe property to guide what and how neural networks learn. The method contains a dual-level domain adaptation framework that synergistically integrates pixel-level and feature-level domain adaptation through joint optimization. Pixel-level domain adaptation enhances visual realism of simulated interferograms. Our feature-level domain adaptation method that leverages fringe semantics depends mainly on spatial structures. It guides the feature processing module to focus on structural patterns while de-emphasizing domain-sensitive distractors. Therefore, the neural network trained with this method will achieve enhanced cross-domain robustness. We also proposed a fringe-context-aware loss function that embeds the fringe continuity property to enhance neural network performance. This integrated approach achieves state-of-the-art segmentation using only 60 unlabeled real interferograms and 30 background images. Our method delivers an annotation-efficient solution for interferogram segmentation, offering actionable insights for deep learning optical image processing under domain shifts and label scarcity.