Geometric-constraint-aware binocular framework for robust absolute phase recovery in structured light 3D...

Area of Science:

• Computer Vision
• Optical Metrology
• Machine Learning

Background:

• Structured light 3D reconstruction faces challenges in accuracy, efficiency, and handling complex surfaces.
• Existing methods struggle with discontinuous surfaces and phase jumps, limiting practical applications.

Purpose of the Study:

• To develop a robust binocular deep learning framework for optical 3D reconstruction.
• To synergistically integrate geometric constraints with data-driven learning for absolute phase recovery.
• To overcome limitations in accuracy, computational efficiency, and robustness to complex surfaces.

Main Methods:

• A novel dual-view structured light system and the StereoPhase dataset with 23,000 synchronized phase maps.
• GCANet, a dual-path network architecture with cross-view feature fusion and dilated residual blocks.
• A consistency-aware loss function incorporating epipolar geometric constraints.

Main Results:

• Achieved competitive speed and accuracy, outperforming existing learning-based approaches on challenging datasets.
• Demonstrated robust absolute phase recovery, effectively addressing phase jumps and discontinuous surfaces.
• Maintained real-time processing capabilities while significantly reducing reconstruction errors.

Conclusions:

• The proposed framework establishes a new paradigm for phase unwrapping, bridging geometric and deep learning methods.
• Offers a powerful solution for industrial inspection and biomedical applications demanding precision and efficiency.
• Successfully balances reconstruction accuracy, computational efficiency, and robustness for structured light 3D systems.