Data-Driven Complementation Approach for High-Reflective Surface Measurement

Optical approaches are widely used in surface geometry measurement for their efficiency and accuracy. However, the imaging quality is strongly affected by the surface reflectance, resulting in incomplete 3D reconstruction of the aim surface. This paper designs a data-driven complementation architecture for high-reflective surface measurement to fix the incomplete reconstruction result. Firstly, we design a dataset based on Bidirectional Reflectance Distribution Function (BRDF) to synthesis the imaging process. Then, we design a non-local inpainting neural network combined the attention architecture to complete the regions that are failed for reconstruction. Experiments show that the system in the paper provides impressive accuracy compared with interpolation-based and convolutional network-based complementation method.