Fringe-prior-guided end-to-end fringe projection profilometry network for textured objects

Fringe projection profilometry (FPP) is a widely used 3D reconstruction method due to its high accuracy and non-contact nature. However, when measuring textured objects, conventional single-shot FPP methods often suffer from reduced 3D reconstruction accuracy and cannot reconstruct the surface texture. To address this challenge, this paper proposes a fringe-prior-guided end-to-end deep learning network that takes a single modulated fringe image and a pre-captured reference fringe image as input and simultaneously achieves high-precision 3D reconstruction and texture reconstruction. To reduce the influence of texture on 3D reconstruction, a Fringe Feature Alignment Module is designed to align the reference fringe with the modulated fringe, thereby generating a relatively clean fringe prior image, a texture prior image, and a confidence map for reliability assessment. To fully exploit these priors for reliable 3D reconstruction and texture reconstruction, a Confidence-guided Reconstruction Backbone is further designed to separate fringe and texture components and reconstruct a high-precision depth map and initial texture map. To further enhance texture details and realism, a Texture Refinement and Compensation Module is designed to extract reliable texture details to refine the initial texture image, yielding the final reconstructed texture. In addition, this paper constructs and publicly releases a synthetic FPP dataset for textured objects using 3ds Max virtual software. Both simulation and physical experiments demonstrate that the proposed method outperforms traditional methods and existing deep learning-based FPP methods in both 3D reconstruction and texture reconstruction accuracy.