Color crosstalk removal based on deep learning for single-shot color FPP method

Lianpo Wang, Yanyang Xing,
Color crosstalk removal based on deep learning for single-shot color FPP method,
Optics and Lasers in Engineering,
Volume 194,
2025,
109211,
ISSN 0143-8166,
https://doi.org/10.1016/j.optlaseng.2025.109211.
(https://www.sciencedirect.com/science/article/pii/S0143816625003963)
Abstract: Single-shot fringe projection profilometry (FPP) is more suitable for 3D measurement of fast-moving dynamic objects, as it only requires one exposure to reconstruct the three-dimensional morphology. However, phase unwrapping based on a single-frame fringe image in discontinuous scenes is ill-posed. Therefore, the single-shot color FPP (SCFPP) method, which use three channels of a color camera to capture three fringe images in one-shot, overcomes the problem of ill-posedness and has been widely studied. Especially, with the introduction of deep learning, SCFPP has even achieved 3D reconstruction results comparable to the multi-shot FPP methods. Nonetheless, the issue of color crosstalk among the three channels of a color camera can seriously affect the accuracy of SCFPP and remains an open problem. To address this issue, we propose an end-to-end SCFPP network model and construct a color fringe simulation dataset with color crosstalk. Specifically, we add three different attention mechanism methods on U-Net, namely Multi-axis External Weight (MEW), Criss-Cross Attention (CCA), and Attention Gate (AG), to filter out irrelevant crosstalk information in channels, spaces, and contexts, respectively. Simulation and real experimental results show that our network outperforms traditional and other deep learning methods. The code and dataset are available at: https://github.com/LianpoWang/CAMEW-UNet.
Keywords: Single-shot fringe projection profilometry; Color fringe projection profilometry; Deep learning; Color crosstalk