光学
计算机科学
投影(关系代数)
结构光三维扫描仪
计算机视觉
轮廓仪
人工智能
物理
噪音(视频)
校准
图像处理
作者
Weikang Chen,Yuxia Li,Xiangting Zhang,X Wang,Xinghui Li
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-04-24
卷期号:: 1022-1028
标识
DOI:10.1109/eicct69950.2026.11564530
摘要
Deep learning-based fringe projection profilometry (FPP) faces critical bottlenecks in industrial deployment, including prohibitive real dataset acquisition costs, poor adaptability of existing simulation data to metal surfaces, and severe simulation-to-real (sim2real) domain shift. This paper presents RFAD, a reprojection-enhanced automated digital twin system built on the Blender Cycles physically based rendering engine. RFAD achieves 1:1 high-fidelity digital twinning of the complete FPP projection-imaging pipeline, supporting precise full-parameter mapping, industrial metal material simulation, and automated batch dataset generation. To bridge the sim2real gap in FPP without breaking phase consistency, we propose a physics-grounded reprojection method. Domain shift is confined to three interpretable parameters-background, modulation, and phase. Using a white flat plate as a domain anchor, we learn per-parameter mappings from paired real-simulated data and reproject them to any complex scene. This achieves phase-preserving, registration-free domain adaptation without black-box style transfer. Extensive verification demonstrates that the enhanced fringe data achieves 99.15% global distribution coincidence with real data and standard sphere reconstruction error within 10 μm in system calibration cross-validation. The deep learning experiment shows that pre-training on simulation data can reduce the phase MAE by 43.8%. RFAD effectively breaks the data bottleneck for industrial FPP and provides a high-fidelity, low-cost solution for data-driven 3D measurement algorithm development.
科研通智能强力驱动
Strongly Powered by AbleSci AI