End-to-end single-shot fringe projection profilometry based on semi-supervised learning

Abstract

Single-shot Fringe Projection Profilometry (SSFPP) enables three-dimensional (3D) reconstruction from a single captured fringe image and is widely applied to dynamic object measurement. Although deep learning has significantly improved the accuracy and robustness of SSFPP, most existing approaches remain fully supervised and rely heavily on labeled data that are difficult and costly to obtain. Existing self-supervised FPP methods circumvent the need for labeled data but typically rely on multiple input images to resolve the inherent $2k\pi$ phase ambiguity of a single fringe pattern. Consequently, they cannot achieve true single-shot FPP reconstruction. To enable self-supervised SSFPP, we propose a dual-domain self-supervised loss function. In the image domain, a Structural Similarity Index Measure (SSIM) loss is introduced to enforce physically meaningful consistency between the reprojected and input fringe images, thereby supporting end-to-end self-supervised learning. In the phase domain, an edge-aware self-smoothing loss is developed to suppress discontinuities caused by the $2k\pi$ phase ambiguity, enabling a unique and spatially continuous phase solution from a single frame. In addition, we design a dynamic dual-stream sampler that simultaneously samples labeled and unlabeled data and adaptively adjusts their proportions within each batch based on training progress, enabling progressive and synergistic optimization of supervised and self-supervised learning signals. Experimental results demonstrate that the proposed method, using only 50% of the labeled data, outperforms existing open-source supervised end-to-end SSFPP approaches on both synthetic and real-world datasets. This confirms its ability to substantially reduce annotation costs while maintaining high reconstruction accuracy.