Untrained deep learning-based phase retrieval for fringe projection profilometry

Abstract

Fringe projection profilometry (FPP) based on deep learning shows potential for challenging 3-D sensing tasks, e.g., bio-medicine, reverse engineering, and face recognition, etc. Supervised deep learning has been introduced to retrieve the desired phase for the 3-D reconstruction, which relies on lots of advanced training to construct the fringe-to-phase transformation. The traditional deep learning-based method becomes unreliable for scenes that are different from the training ones, which restricts it to be applied for actual applications. In this paper, an untrained deep learning-based phase retrieval method is proposed. By adding a camera to the traditional FPP system, scene-independent physical constraints such as phase, structure and 3-D consistency are constructed to optimize the fringe-to-phase transformation. The proposed deep learning-based method can retrieve the desired phase from a single fringe pattern without advance training. Both theoretical analyses and experimental results demonstrate its accurateness and robustness. The proposed method also shows potential for single-shot 3-D sensing applications such as high-speed 3-D measurement.