3D shape measurement based on Res-Attention-Unet for deep learning

This paper proposes a 3D shape measurement method based on Res-Attention-Unet for deep learning, which effectively integrates single-frame fringe phase extraction and temporal phase unwrapping. The Res-Attention-Unet, trained with a large amount of data, is shown to be able to extract phase information directly from deformed fringe patterns with large frequency differences and to achieve high-precision and unambiguous 3D measurement using hierarchical phase unwrapping. We experimentally demonstrate that the method requires only three deformed fringe patterns with different frequencies to achieve accurate and 3D measurement, and Res-Attention-Unet has an absolute advantage over the traditional single-frame fringe phase extraction algorithm in terms of the phase extraction accuracy of low-frequency fringe patterns. Since only three fringe patterns need to be projected, this method is expected to be applied to project single-frame color fringe for dynamic 3D measurement.