Deep Learning-Based Phase Recovery and Image Analysis for Laser Interferometric Measurements

Laser interferometry is widely used to evaluate the surface errors in low and middle frequency domains. Despite its high theoretical accuracy, the measurement result is influenced by phase ambiguity and noise sensitivity in practice. This study proposes a three-stage framework integrating physical modelling, intelligent computation, and 3D defect detection to enhance measurement reliability by deep learning. The framework begins with the physical model of interferometry. By incorporating Zernike polynomials to simulate high-order aberrations and adding stochastic noise, a synthetic fringe pattern dataset is generated to mimic the errors from the surface and measurement process. Continuous phase maps are extracted via path-dependent phase unwrapping algorithms as ground-truth labels, enabling the training of a residual-augmented ResU-Net architecture. The network’s skip connections preserve high-frequency details while residual blocks mitigate noise effects, achieving robust phase demodulation. The 3D surface topography is finally reconstructed, and a cross-module YOLOv11 architecture synchronously analyzes 2D phase gradients and 3D point cloud curvature. This dual-domain approach enables automated defect localization through adaptive feature fusion. It is demonstrated that the surface discontinuities can be detected without manual parameter tuning.