Directional periodicity-aware factorized attention network for single-shot fringe projection profilometry.

Area of Science:

• Computer Vision
• Metrology
• Machine Learning

Background:

• Single-shot fringe projection profilometry is crucial for real-time 3D measurement.
• Existing deep learning methods struggle with degraded fringe patterns, including sensor noise and shadows.
• Current approaches often indiscriminately process reliable and corrupted image regions, leading to inaccurate depth estimation.

Purpose of the Study:

• To propose a novel deep learning framework, FPDNet, for robust single-shot fringe projection profilometry.
• To address challenges posed by noise and shadow regions in practical fringe patterns.
• To enhance the accuracy and reliability of 3D measurements in real-world scenarios.

Main Methods:

• Developed FPDNet, a deep learning framework incorporating a pixel-wise reliability attention module (PRAM) and a factorized efficient spatial attention (FESA) module.
• PRAM utilizes local statistics to generate pixel-wise reliability maps, enabling adaptive feature fusion.
• FESA employs factorized attention with elongated kernels to exploit fringe pattern periodicity and suppress noise.

Main Results:

• FPDNet achieved a mean absolute error (MAE) of 1.339 mm under low-noise conditions, a 54.7% reduction compared to the baseline.
• Demonstrated superior noise robustness, with only a 12.3% increase in MAE from low to high noise levels.
• Maintained real-time processing speeds of 52 FPS.

Conclusions:

• FPDNet effectively mitigates the impact of noise and shadows in fringe projection profilometry.
• The proposed physics-guided attention modules significantly improve depth estimation accuracy and robustness.
• FPDNet represents a significant advancement for real-time, high-accuracy 3D measurement systems.