Computational Imaging for Accurate 3D Modeling in Diverse Fields

Xu, Bingjie (Jenny)

doi:https://doi.org/10.21985/n2-tf3g-zf62

Work

Computational Imaging for Accurate 3D Modeling in Diverse Fields

Public

Download PDF

Download All Files (.zip)

This dissertation introduces several novel computational imaging techniques that capture and analyze the 3D surface shapes and internal layered materials. The research proposes user-friendly and non-invasive imaging systems, constructed using only commercial off-the-shelf (COTS) components, which provide accurate measurement of 3D information that was previously inaccessible. The dissertation focuses on two areas of 3D modeling: surface and volumetric modeling. For surface modeling, the research introduces a mobile photometric stereo system called ‘shape-from-shifting’, and two mobile phase measuring deflectometry systems that capture surface features of large-scale mostly-flat objects and mostly-spherical objects. It showcases their implementation with cultural heritage applications, by detecting protrusions in oil paintings, determining the provenance of stained glass pieces, and by tracking eye movements for virtual reality (VR) headsets. In comparison to previous studies that mostly relied on bulky and lab-specific systems, our imaging systems use a mobile device screen or a wall, and commercial camera(s) or a projector, accompanied by open-source Python packages and tutorials, enabling flexible usage and accessibility to non-experts. For volumetric modeling, I present a novel time-domain optical coherence tomography (OCT) system that uses off-the-shelf optical components and a dynamic-focus approach to compensate for low-energy problems in commercial low-cost continuum laser sources. This system can penetrate multi-layer oil paintings with IR wavelengths that reach up to 2 μm, making it the first low-cost time-domain OCT system at 2 μm for cultural heritage applications. To analyze the chemical composition of multi-layer oil paintings, the research proposes a deep-learning framework that uses X-Ray Fluorescence (XRF) spectra, enabling automatic identification of mixed pigments for multi-layer oil paintings, which is the first study to use deep learning for this purpose.

Creator