Neural measurement method based on near-field photometric stereo with tangent estimation for machined metal surfaces

Abstract

Machined metal surfaces exhibit non-Lambertian reflectance with complex highlight that dramatically undermines the performance of three-dimensional optical measurement methods. To address this problem, this study proposes a neural measurement method based on near-field photometric stereo to obtain accurate measurements of machined metal surfaces. A tangent estimation network is designed to enable the photometric stereo to suppress the effect of the anisotropic metal reflectance on the surface normal estimation. In addition, a projector is introduced into the photometric stereo system to output an initial point cloud that helps the network adapt to near-field scenes and rectify the error accumulation of normal integration. Synthetic experiments based on public anisotropic reflectance data demonstrate that the proposed photometric stereo network with tangent estimation improves the surface normal estimation accuracy of anisotropic metal reflectance surfaces. Real experiments based on eight machined metal surfaces demonstrate that the proposed neural measurement method obtains a measurement accuracy of 59 um compared with the measurement results of the coordinate measurement machines.