Yao Pengcheng, Gai Shaoyan, Da Feipeng
Opt Lett. 2021 Sep 15;46(18):4442-4445. doi: 10.1364/OL.431676.
Fringe projection profilometry (FPP) is one of the most widely used 3D reconstruction techniques. A higher-resolution fringe pattern produces a more detailed and accurate 3D point cloud, which is critical for 3D sensing. However, there is no effective way to achieve FPP super-resolution except by using greater hardware. Therefore, this Letter proposes a dual-dense block super-resolution network (DdBSRN) to extend the fringe resolution and reconstruct a high-definition 3D shape. Especially, a novel dual-dense block structure is designed and embedded into a multi-path structure to fully utilize the local layers and fuse multiple discrete sinusoidal signals. Furthermore, a fully functional DdBSRN can be obtained even when training with a smaller data sample. Experiments demonstrate that the proposed DdBSRN method is stable and robust, and that it outperforms standard interpolation methods in terms of accuracy and 3D details.
条纹投影轮廓术（FPP）是应用最为广泛的三维重建技术之一。更高分辨率的条纹图案能生成更详细、准确的三维点云，这对三维传感至关重要。然而，除了使用更强大的硬件外，尚无实现FPP超分辨率的有效方法。因此，本文提出一种双密集块超分辨率网络（DdBSRN）来扩展条纹分辨率并重建高清三维形状。特别地，设计了一种新颖的双密集块结构并将其嵌入多路径结构中，以充分利用局部层并融合多个离散正弦信号。此外，即使使用较小的数据样本进行训练，也能获得功能完备的DdBSRN。实验表明，所提出的DdBSRN方法稳定且鲁棒，在精度和三维细节方面优于标准插值方法。
