Enhancing 3D scanning of reflective metal surfaces via pseudo-exposure and retinex-based deep learning

Structured-light scanning of highly reflective metal surfaces often suffers from pixel saturation, leading to blurred or lost fringes that hinder 3D reconstruction accuracy. This study introduces a system integrating pseudo-exposure technology and a Retinex-inspired deep learning network ( https://github.com/tankKuo2003/Retinex-Low-Light-Enhancement-TF2 ) to address this challenge. By generating multiple virtual exposures through gamma mapping, the system reduces capture time and enhances low-exposure images, revealing details in shadowed regions. A multi-exposure pyramid image fusion technique then merges optimal pixels, producing high dynamic range images with clear fringes. Experimental results on flat and cylindrical metallic workpieces demonstrate significant improvements in point cloud reconstruction accuracy and completeness.