Adaptive Structured-Light 3D Surface Imaging with Cross-Domain Learning

Abstract

The rapid development of artificial intelligence (AI) technology is leading a paradigm shift in optical metrology, from physics- and knowledge-based modeling to data-driven learning. In particular, the integration of structured-light techniques with deep learning has garnered widespread attention and achieved significant success due to its capability to enable single-frame, high-speed, and high-accuracy 3D surface imaging. However, most algorithms based on deep neural networks (DNNs) face a critical challenge: they assume the training and test data are independent and identically distributed, leading to performance degradation when applied across different image domains, especially when test images are acquired from unseen systems and environments. A cross-domain learning framework for adaptive structured-light 3D imaging is proposed to address this challenge. This framework's adaptability is enhanced by a novel mixture-of-experts (MoE) architecture, capable of dynamically synthesizing a network by integrating contributions from multiple expert DNNs. Experimental results demonstrate the method exhibits superior generalization performance across diverse systems and environments over both “specialist” DNNs developed for fixed domains and “generalist” DNNs trained by brute-force approaches. This work offers fresh insights into substantially enhancing the generalization of deep-learning-based structured-light 3D imaging and advances the development of versatile, robust AI-driven optical metrology techniques.