人工智能
计算机科学
计算机视觉
斑点图案
稳健性(进化)
轮廓仪
三维重建
数字图像相关
结构光
极线几何
迭代重建
投影机
匹配(统计)
曲面重建
校准
曲面(拓扑)
投影(关系代数)
立体视觉
结构光三维扫描仪
双向反射分布函数
图像(数学)
束流调整
立体成像
算法
数字光处理
深度学习
数字图像
杂乱
图像处理
作者
Zizhan Gao,Yanzhao Liu,Jianhui Zhao,Bing Pan
标识
DOI:10.1088/1361-6501/aebb50
摘要
Existing speckle pattern profilometry (SPP) methods for 3D shape measurement generally adopt subset/template-based image matching algorithms for disparity estimation. Despite their attractive merit in single-shot 3D reconstruction, these methods are computationally intensive and often suffer from data loss in regions with sharp gray-level variations or complex surface geometries. To overcome these limitations, an improved SPP using a novel deep-learning based pixelwise image registration framework is introduced. Specifically, a SPP system utilizing a calibrated stereovision setup integrated with a digital projector is established to capture projected speckle patterns. Disparity between the stereo pair that encodes the surface topography is estimated using a state-of-the-art deep learning-based digital image correlation (DL-DIC) framework, called G-RAFT. The estimated disparity is subsequently combined with stereo calibration parameters for triangulation-based 3D reconstruction. By using G-RAFT, the proposed DL-DIC-aided SPP achieves efficient, single-shot, high-accuracy, and dense 3D shape measurement. Experimental results on standard geometries demonstrate that the proposed method achieves reconstruction accuracy comparable to traditional subset/template-based algorithm (Smart-DIC) while providing higher computational efficiency. Experiments on a complex plaster bust further show that this DL-DIC-aided SPP improves reconstruction completeness and robustness in regions where traditional subset-based matching exhibits data loss or mismatches. These results demonstrate that the proposed DL-DIC-aided SPP enables accurate, robust, and dense single-shot 3D reconstruction while preserving the explicit calibration-based geometric reconstruction pipeline.
科研通智能强力驱动
Strongly Powered by AbleSci AI