Deep Learning Enhanced Dynamic 3-Dimensional Shape Measurement Using Single-Shot Spatial Multiplexing and Transformer-Based Phase Retrieval

This is the authors' abstract. We don't add key points for this paper.

Real-time, high-precision 3-dimensional (3D) imaging is essential for applications such as industrial inspection, robotic navigation, and human–computer interaction. Fringe projection profilometry (FPP), a widely used structured light method, achieves high spatiotemporal resolution by rapidly projecting and processing of fringe patterns. However, traditional multi-frame FPP methods are hindered by motion-induced artifacts and computational bottlenecks, limiting their applicability in dynamic environments. In this work, we propose a multiplexed structured light 3D measurement method that integrates a physics-based Transformer framework to minimize the number of projection patterns required for precise single-snapshot measurements. This method extracts accurate phase information from a single fringe image, enabling artifact-free, high-resolution 3D surface reconstruction. By combining low-frequency triangular waves with high-frequency sinusoidal fringes, we ensure unambiguous phase retrieval, providing the deep neural network with reliable inputs. The Transformer-based network leverages superior global information capture and multi-scale feature learning capabilities for robust fringe analysis and phase unwrapping, thereby enhancing the accuracy and generalization of depth prediction. Experimental evaluations demonstrate that our method outperforms traditional single-frame phase retrieval techniques and other deep learning-based methods in terms of precision and robustness. Dynamic measurements of complex objects with various materials further validate its potential for high-speed, real-time 3D imaging in intelligent manufacturing and augmented reality.