Large-Area Reflection Suppression for Structured-Light Three-Dimensional Reconstruction of Specular Surfaces

Abstract

The inherent physical constraints of polishing tools lead to positioning and repeatability errors in industrial robots, which compromise polishing accuracy by deviating the actual dwell time from its theoretical value. A structured light detection technology, as a high-precision non-contact measurement technology, provides a promising solution for the on-machine positioning of polished workpieces. Nevertheless, the strong specular reflection of metal workpieces tends to cause large area overexposure in the captured fringe images, thereby inducing the loss of three-dimensional point cloud data. An algorithm combining the improved Retinal-cortex theory (Retinex) and Mask-Aware Transformer (MAT) deep learning is proposed to suppress the influence of large area surface reflection on measurement to address this issue in this study. The process involves adaptive aperture adjustment to obtain images within an optimal grayscale range, followed by Retinex-based stronglight suppression and MAT-based inpainting to recover missing fringe information in overexposed areas. Validation experiments are conducted on high-reflective aluminum alloy plates based on a structured light measurement system calibrated with an accuracy of 0.01 mm. The results demonstrate that the proposed algorithm can effectively recover missing surface-data holes in the point cloud with an average growth rate of 3.92%, eliminate overexposed regions and compensate for underexposed areas, providing a novel solution for the on-machine inspection and localization of polished workpieces. It is of great significance for improving the deterministic polishing accuracy of robots.