Coherent and Holographic Imaging Methods for Immersive Near-Eye Displays

Near-eye displays have been designed to provide realistic 3D viewing experience, strongly demanded in applications, such as remote machine operation, entertainment, and 3D design. However, contemporary near-eye displays still generate conflicting visual cues which degrade the immersive experience and hinders their comfortable use. Approaches using coherent, e.g., laser light for display illumination have been considered prominent for tackling the current near-eye display deficiencies. Coherent illumination enables holographic imaging whereas holographic displays are expected to accurately recreate the true light waves of a desired 3D scene. However, the use of coherent light for driving displays introduces additional high contrast noise in the form of speckle patterns, which has to be taken care of. Furthermore, imaging methods for holographic displays are computationally demanding and impose new challenges in analysis, speckle noise and light modelling.

This thesis examines computational methods for near-eye displays in the coherent imaging regime using signal processing, machine learning, and geometrical (ray) and physical (wave) optics modeling. In the first part of the thesis, we concentrate on analysis of holographic imaging modalities and develop corresponding computational methods. To tackle the high computational demands of holography, we adopt holographic stereograms as an approximative holographic data representation. We address the visual correctness of such representation by developing a framework for analyzing the accuracy of accommodation visual cues provided by a holographic stereogram in relation to its design parameters. Additionally, we propose a signal processing solution for speckle noise reduction to overcome existing issues in light modelling causing visual artefacts. We also develop a novel holographic imaging method to accurately model lighting effects in challenging conditions, such as mirror reflections.

In the second part of the thesis, we approach the computational complexity aspects of coherent display imaging through deep learning. We develop a coherent accommodation-invariant near-eye display framework to jointly optimize static display optics and a display image pre-processing network. Finally, we accelerate the corresponding novel holographic imaging method via deep learning aimed at real-time applications. This includes developing an efficient procedure for generating functional random 3D scenes for forming a large synthetic data set of multiperspective images, and training a neural network to approximate the holographic imaging method under the real-time processing constraints.

Altogether, the methods developed in this thesis are shown to be highly competitive with the state-of-the-art computational methods for coherent-light near-eye displays. The results of the work demonstrate two alternative approaches for resolving the existing near-eye display problems of conflicting visual cues using either static or dynamic optics and computational methods suitable for real-time use. The presented results are therefore instrumental for the next-generation immersive near-eye displays.