A multi-code 3D measurement technique based on deep learning

Benefiting from the merits of low cost, high accuracy and high resolution, fringe projection profilometry has been developing rapidly over the past decades. However, recovering the absolute phase with high accuracy and robustness effectively has always been significant challenge in fringe projection profilometry. In this paper, an intelligent Multi-code Deep Learning (MCDL) technique is developed to solve the high-slope absolute phase from only two patterns with high accuracy and robustness. Two sub-networks are designed for obtaining the wrapped phase and the fringe order. Specially, the proposed MCDL method can solve the high-level fringe orders by only a special multi-code pattern itself through a cooperative multi-connected convolutional neural network. By training the deep network with numerous datasets, the principle of unwrapping phase can be learned by the MCDL approach. Experiments demonstrate that the proposed method has the abilities of high robustness, efficiency and accuracy (measurement error: 0.0189mm, FOV: 250 mm x 200 mm), indicating potential applications for highspeed and high-accuracy three dimensional optical measurement.