Single-shot high-precision 3D reconstruction with color fringe projection profilometry based BP neural network

Abstract

In order to reduce the influence of nonlinear factors such as gamma nonlinearity and color crosstalk on the phase retrieval accuracy with color fringe projection profilometry (CFPP) in rapid measurement, a single-shot high-precision backpropagation (BP) neural network based CFPP is proposed. Firstly, neural networks are trained from samples through a training process. The neural network inputs are the RGB components of composite color fringe patterns. The numerator and denominator of the arctangent function are used as network outputs. The monochromatic fringe patterns are adopted. The numerator and denominator of the arctangent function are calculated by the 15-step phase-shift method, which serves as the network training target. Applying this neural network to the prediction of the standard phase can reduce the influence of nonlinear factors on the phase accuracy. Then, the absolute phase was unwrapped by using geometric constraints. Finally, the 3D reconstruction with high precision can be realized by using only a single-shot color fringe. The theoretical model of the proposed method and the process of neural network training are introduced. The results of 3D reconstruction are compared with other algorithms in the same condition. The experimental results demonstrate that the RMS of the proposed algorithm can reach 0.0107 mm, compared with 0.0121 mm of the traditional 3-step phase-shift method. The proposed method is applied to the online quality inspection of relay components, and the results verify the high accuracy and high efficiency of measurement.