An End-to-End Network for 3D Reconstruction from Multi-Frequency Fringe Patterns

Fringe Projection Profilometry (FPP), characterized by its non-invasive nature and fine measurement accuracy, is extensively utilized in precise 3D shape acquisition. However, conventional FPP frameworks often rely on sequential procedures-such as phase retrieval and depth estimation-which tend to introduce cumulative errors and hinder the overall efficiency of the reconstruction process. To address this, we propose F2PNet, an end-to-end neural network that directly reconstructs 3D point clouds from multi-frequency fringe images. Unlike traditional approaches dependent on intermediate phase retrieval, F2PNet directly infers the full 3D coordinates (X, Y, Z) from four input fringe patterns of distinct frequencies. The proposed architecture is built upon a U-Net architecture and integrates ResNet blocks to enhance feature representation. Experimental results demonstrate that F2PNet achieves superior reconstruction accuracy and generalization performance compared to existing baseline models, across both complex object surfaces and standard spheres.