Optical 3-D Measurement for Low-SNR Scenes via Physics-Informed Zero-Shot Learning

In industrial 3D metrology, the low signal-to-noise ratio (SNR) issue is commonly encountered, due to inappropriate illumination intensity, limited imaging dynamic range, or complex scene material, etc. Compared with non-learning-based methods, deep-learning-based methods excel in efficiency and fidelity for the low SNR issue. However, most of them are data-driven, thus have limited generalization ability. Besides, they require advanced computing hardware for network training, greatly increasing the metrology cost. To tackle these problems, a physics-informed zero-shot learning (PZL) method with an ultra-lightweight neural network (UNN) is proposed for low-SNR scene measurement. There are two major contributions in our method. First, by blending physics priors for phase retrieval and fringe noise, a generalized PZL framework with a noisy-sinusoidal-component-to-noisy-sinusoidal-component mapping is established. The low SNR issue of various challenging scenes including the low-illumination, high-dynamic-range, strong-ambient-light, and large-depth-range scenes is unified in a single enhancement framework. Moreover, no training dataset is required other than the degraded fringe itself, and the generalization ability for fringe enhancement is significantly improved. Second, based on the PZL framework, a symmetrized optimization strategy along with the UNN is proposed. Valid 3D reconstruction of fine surface details can be achieved on computing-resource-constrained platforms, even on a CPU. Experiments verify the superiority of our method in efficiency, fidelity, generalization ability and computing hardware cost. And to our knowledge, it is the first time such a simultaneous achievement has been accomplished.