Fringe property-guided deep learning method for interferogram segmentation.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Accurate interferogram segmentation is vital for optical measurements and metrology.
• Deep learning for interferogram segmentation faces challenges due to limited annotated real interferograms and domain gaps between simulated and real data.

Purpose of the Study:

• To develop an annotation-efficient deep learning method for interferogram segmentation that bridges the domain gap between simulated and real data.
• To enhance the cross-domain robustness of neural networks for optical image processing.

Main Methods:

• Proposed a fringe property-guided deep learning method incorporating a dual-level domain adaptation framework (pixel-level and feature-level).
• Implemented feature-level domain adaptation leveraging fringe semantics and spatial structures to focus on structural patterns.
• Introduced a fringe-context-aware loss function embedding fringe continuity property.

Main Results:

• Achieved state-of-the-art segmentation performance using only 60 unlabeled real interferograms and 30 background images.
• Demonstrated enhanced cross-domain robustness by guiding neural network learning with fringe properties.
• The dual-level domain adaptation synergistically improved visual realism and feature focus.

Conclusions:

• The proposed method offers an annotation-efficient solution for interferogram segmentation, crucial for optical metrology.
• Provides actionable insights for deep learning in optical image processing facing domain shifts and label scarcity.
• Highlights the effectiveness of incorporating domain knowledge (fringe properties) into deep learning models.