Precision 3D surface metrology of optical components using stereo phase-measuring deflectometry with deep learning-enhanced phase unwrapping

We present a stereo phase-measuring deflectometry (PMD) system for precision inspection of specular optics. The pipeline combines (i) a geometry module that recovers the screen-camera extrinsics from dense mappings between image phase and LCD coordinates, initialized with a virtual-camera reflection seed and refined via joint stereo bundle adjustment, and (ii) a lightweight attention U-Net that unwraps horizontal/vertical phases jointly using a Charbonnier+RMS+gradient loss to preserve continuity while resisting noise. Using two 4708×2824 area-scan cameras and a 4K monitor, the calibrated stereo model attains an overall mean reprojection error of 0.189 pixels. On a 50×50 mm protected-silver flat mirror, the baseline stereo reconstruction without deep unwrapping yields PV/RMS residuals of 1575.2 nm/562.9 nm. Replacing the phase unwrapping with our network reduces the residuals to 1096.4 nm/385.6 nm, i.e., 30.4% (PV) and 31.5% (RMS) improvements under identical geometry. Qualitatively, the binocular fusion suppresses edge drift seen in monocular integration, while the learned unwrapping mitigates low-frequency bias, leading to micrometer-scale height consistency across the aperture. These results demonstrate that coupling stereo geometric constraints with deep phase processing provides a practical accuracy gain for high-resolution PMD of optical components within an efficient, all-digital setup.