Three-dimensional measurement for specular reflection surface based on deep learning and phase measuring profilometry

Abstract

ObjectivePhase measuring profilometry (PMP) is a non-contact 3D measurement method. However，it is still a very challenging task when it is applied to an object with a specular reflection surface. In this paper, a 3D measurement method for specular reflection surface based on a deep learning framework is proposed to reduce the measurement error caused by highlights. MethodsFirstly, a dual-channel U-Net network (DC-UNet) is used to identify the highlight areas in the fringe image. Secondly, the dilated residual networks (DRN) combined with the features of the neighborhood of the highlight pixels are used to correct the highlight pixels, and the predicted fringe image without highlight can be obtained. Thirdly, the predicted fringe image is blended with the original image, regarded as the final result of the highlight removal. Fourthly, phase unwrapping is implemented using the fringe image of highlight removal, and the 3D coordinates values can be obtained by combining the calibrated parameters with the phase values unwrapped. ResultsSimulated results show that the absolute maximum deviation, absolute mean error and standard deviation of reconstructed 3D height coordinates are reduced by 98.27 %, 68.70 %, and 91.11 %, respectively. The real experiment results show that the maximum error of the absolute phase is reduced by more than 87.00 %, and the mean error of the absolute phase is reduced by more than 92.72 %. Finally, more than 99.14 % of the pixels with intensity saturated caused by highlights can be corrected by the proposed method.