Deep-learning-enabled single-shot FPP using spatially composite fringes and dual fringe-order fusion

Single-shot fringe projection profilometry (FPP) has received extensive attention due to its efficiency in dynamic 3D measurement, but existing deep learning methods often suffer from fringe order misjudgment and phase unwrapping errors on complex or isolated surfaces due to the coupling between wrapped phase and fringe order and the conflict between regression and classification tasks. To address these issues, we propose an improved inner shifting-phase spatial composite FPP method combining spatial separation composite fringe coding with internal phase shift level embedding. The method spatially composites multifrequency fringes and embeds fringe order information as an internal phase shift into the high- frequency component, physically decoupling phase and order. A multitask U-Net with a shared shallow encoder and two task-specific decoder branches is built: the phase decoder regresses unit-circle representations of the three frequency wrapped phases, whereas the order decoder classifies the embedded fringe orders. We further introduce an absolute phase retrieval strategy that fuses physical consistency derived from multifrequency heterodyne with the learned local order and spatial continuity, achieving robust phase recovery. Experimental results demonstrate high accuracy and stable reconstruction on complex surfaces and isolated objects, competitive with conventional multiframe FPP, and dynamic experiments confirm its potential for high-speed 3D measurement.