LiteF2DNet: a lightweight learning framework for 3D reconstruction using fringe projection profilometry

Fringe projection profilometry (FPP) is the most commonly used structured light approach for 3D object profiling. Traditional FPP algorithms have multistage procedures that can lead to error propagation. Deep-learning-based end-to-end models currently have been developed to mitigate this error propagation and provide faithful reconstruction. In this paper, we propose LiteF2DNet, a lightweight deep-learning framework to estimate the depth profile of objects, given reference and deformed fringes. The proposed framework has dense connections in the feature extraction module to aid better information flow. The parameters in the framework are 40% less than those in the base model, which also means less inference time and limited memory requirements, making it suitable for real-time 3D reconstruction. To circumvent the tedious process of collecting real samples, synthetic sample training was adopted in this work using Gaussian mixture models and computer-aided design objects. The qualitative and quantitative results presented in this work demonstrate that the proposed network performs well compared to other standard methods in the literature. Various analysis plots also illustrate the model's superior performance at high dynamic ranges, even with low-frequency fringes and high noise. Moreover, the reconstruction results on real samples show that the proposed model can predict 3D profiles of real objects with synthetic sample training.