Single-Shot 3D Reconstruction of Microscopic Objects with Structured Light and Fringe-to-Phase Neural Network

Deep learning plays a pivotal role in the field of three-dimensional (3D) optical measurement. This paper proposes a robust single-shot 3D reconstruction method that combines structured light technology with convolutional neural network. The proposed fringe-to-phase network takes a single deformed grayscale fringe pattern as the input and outputs an absolute phase containing object height information, which can be used to obtain the corresponding 3D shape directly from a 2D image of fringe pattern without additional processing. A dataset with high-quality ground truth labels is generated by a phase shifting and multi-frequency heterodyne algorithm, possessing a diverse representation. The phase accuracy of the model output is enhanced through the implementation of hyperparameter optimization, residual block and feature fusion. Accordingly, the number of fringe patterns could be significantly reduced compared to traditional image analysis methods, while maintaining high reconstruction accuracy. Experiments on 3D reconstruction of the ball grid array (BGA) chip and fine ceramic are conducted to verify the effectiveness and accuracy of the proposed method.