High-precision structured light 3D reconstruction of highly reflective objects using deep learning

Structured light technology, which combines the phase-shifting method and Gray code, enables high-precision three-dimensional (3D) reconstruction. However, when measuring highly reflective objects, specular reflections often cause image distortion and reconstruction failures. To address this challenge, this study proposes a deep-learning-based method. First, an image enhancement network is incorporated into the preprocessing stage to improve stripe-detail features and to mitigate the adverse effects of overexposure and underexposure on stripe-image quality. Subsequently, an image restoration network is used to restore distorted images. Additionally, we construct a real-world stripe-pattern dataset specifically collected from highly reflective objects. Experimental comparisons between this approach and existing techniques demonstrate its effectiveness in restoring distorted images, significantly improving the completeness and quality of 3D reconstructions for such objects.