Single Fringe Phase Retrieval for Translucent Object Measurements Using a Deep Convolutional Generative Adversarial Network

Fringe projection profilometry (FPP) is a measurement technique widely used in the field of 3D reconstruction. However, it faces issues of phase shift and reduced fringe modulation depth when measuring translucent objects, leading to decreased measurement accuracy. To reduce the impact of surface scattering effects on the wrapped phase during actual measurement, we propose a single-frame phase retrieval method named GAN-PhaseNet to improve the subsequent measurement accuracy for translucent objects. The network primarily relies on a generative adversarial network framework, with significant enhancements implemented in the generator network, including integrating the U-net++ architecture, Resnet101 as the backbone network for feature extraction, and a multilevel attention module for fully utilizing the high-level features of the source image. The results of the ablation and comparison experiment show that the proposed method has superior phase retrieval results, not only achieving the accuracy of the conventional method on objects with no scattering effect and a slight scattering effect but also obtaining the lowest errors on objects with severe scattering effects when compared with other phase retrieval convolution neural networks (CDLP, Unet-Phase, and DCFPP). Under varying noise levels and fringe frequencies, the proposed method demonstrates excellent robustness and generalization capabilities. It can be applied to computational imaging techniques in the fringe projection field, introducing new ideas for the measurement of translucent objects.