Development of deep learning aided surface reconstruction system for dimensional metrology using sheet of light scanning technique

Abstract Surface reconstruction is essential in manufacturing to ensure dimensional accuracy and aesthetic quality. Sheet of Light (SoL) 3D reconstruction systems, based on structured light principles, offer high-speed, non-contact scanning, making them ideal for industrial inspection and metrology. Unlike conventional 2D imaging, they capture dense point clouds, enabling precise 3D surface measurement. This work develops a Sheet of Light 3D reconstruction system and enhances its accuracy using state-of-the-art point cloud completion networks. A dataset of 10,000 scanned and augmented models with complex geometries and varying sizes is created. The performance of the raw Sheet of Light reconstruction is evaluated using Chamfer Distance (CD), Earth Mover’s Distance (EMD), and measurement uncertainty metrics. Results show that Adaptive Point Transformer (AdaPoinTr), a geometry-aware transformer architecture, significantly improves surface reconstruction, reducing CD and EMD by over 50% compared to the raw output. A measurement uncertainty study further validates its metrological performance, with AdaPoinTr achieving the lowest mean (0.01258) and standard deviation (0.00925). These findings highlight the Sheet of Light system’s potential for precision-driven manufacturing applications while demonstrating that deep learning-aided point cloud completion enhances accuracy, making it a reliable solution for industrial metrology.