Abstract Deep learning has fostered substantial advancements in fringe projection profilometry (FPP), enabling high-speed and high-accuracy reconstruction, even in uncontrolled environments. However, current deep learning models often lack interpretability, and their precision and transferability in FPP remain suboptimal. This review aims to bridge the gap between the potential and practical application of deep learning methods in FPP by comprehensively surveying advancements in deep learning driven single-frame FPP. We establish methodologies for dataset construction and explore state-of-the-art neural networks, including physics-enhanced architectures. To identify optimal models for various FPP tasks, we critically analyze various metrics and benchmarking. We also delve into generalization techniques such as transfer learning and suggest leveraging large vision models to reduce pretraining costs and enhance the performance of FPP systems. By proposing these advancements, this review seeks to enhance the practical adoption of deep learning-driven FPP and pave the way for its transformative applications across industrial, medical, and scientific fields.
Li et al. (Tue,) studied this question.