High precision 3D measurement with few images based on deep learning

Fringe projection profiling (FPP) is a technique to obtain the three-dimensional shape of an object by projecting periodic fringes onto its surface and analyzing the modulated fringes.The goal of this technique is to quickly and accurately obtain the three-dimensional shape of an object with as few fringe patterns as possible. This paper combines the fringe analysis steps of fringe projection profiling and deep learning, the proposed DARUNet network (Dense and Residual U-Net) introduces Dense Block and Residual Block on the basis of U-Net. Only three modulated fringe patterns with different frequencies need to be captured as the input of the DARUNet network, the network outputs the numerator and denominator of the wrapped phase corresponding to each frequency. After some post-processing, the three-dimensional shape of the object can be obtained. Deep learning relies on high-quality datasets, so this paper compares two methods for temporal phase unwrapping: Multi-frequency (hierarchical) and Multi-wavelength (heterodyne).The Multi-frequency method, which demonstrated superior performance, was chosen to create a high-precision 3D measurement dataset. Experiments show that the proposed network has higher precision in predicting the wrapped phase than U-Net and its series networks, and predicting the numerator and denominator of wrapped phase by fringes is also the optimal route for 3D reconstruction technology based on deep learning, this method achieves a high level of precision with a phase error of less than 0.1 radians and a depth error of less than 0.3 mm. Therefore, the method employed in this paper enables high-precision 3D measurements using only three frames of fringe patterns.