Deep learning assisted high-speed fringe projection profilometry

Abstract

Optical systems for 3D reconstruction are essential tools for recovering the topographic information of objects or scenes by employing optical elements such as cameras and projectors. These systems can be classified into active and passive methods. Active methods utilize projection elements to capture topographical details, while passive methods rely solely on capturing systems like cameras to infer 3D information. Among active optical methods, Fringe Projection Profilometry (FPP) is widely recognized for its precision in 3D surface measurement. However, one of the main limitations of FPP lies in the phase unwrapping process, which requires capturing multiple fringe patterns and can be computationally demanding. This thesis proposes a deep learning-assisted approach to improve the speed of phase unwrapping and, consequently, the 3D reconstruction process. The proposed method leverages a transformer-based architecture fine-tuned to predict an initial depth map of the object, significantly reducing the number of captured images required for accurate phase analysis. By integrating this machine learning model, the phase unwrapping process becomes faster and more efficient, enabling real-time or high-speed 3D measurement applications without sacrificing accuracy. Experimental results demonstrate that the deep learning-enhanced FPP reduces the computational time for phase unwrapping while maintaining high reconstruction accuracy. This research advances the capabilities of Fringe Projection Profilometry by introducing a machine learning approach to address the long-standing challenge of phase unwrapping speed, making it a viable solution for high-speed industrial and scientific applications.

Trabajo de grado -- Facultad de Ingeniería