Single-shot phase extraction for fringe projection profilometry using deep convolutional generative adversarial network

This paper presents a single-shot phase extraction approach based on a deep convolutional generative adversarial network that generates a phase map and a quality mask from an input fringe pattern image. A novel loss function is proposed, and a large-scale (28 800 samples) real fringe pattern dataset is collected to train the network. The experiments demonstrate that the proposed method achieves significantly improved phase extraction accuracy and overcomes the main limitations of Fourier transform profilometry. In addition, the proposed method presents excellent performance for real-time computing, reaching approximately 100 f s(-1) with a single GPU. Moreover, the proposed learning-based approach can automatically perform denoising and phase extraction, without any manually set parameters.