ORIGINAL

Abstract

Holography is said to be the most promising technology that could help realise wide field-of-view displays for augmented and virtual reality in a compact form factor such as an ordinary pair of eyeglasses. In this thesis, we explore the work I have done alongside professors and students at Princeton and the University of North Carolina to develop a solution to improve image quality on existing holographic displays. We propose using a deep learning based approach to model the deviations of our holographic display prototype from the ideal forward model. Using this learned forward model, we then compensate for the deviations in the hologram generation process.

As a senior thesis is required to be independent work, this thesis focuses on my contributions to the project which were mostly focused on designing and optimising the deep learning framework used in our hologram optimisation pipeline. Our model needs to learn the mapping from the target image to the aberrated version we see on our hardware. To this end, we employ a conditional generative adversarial network, and pose this as an image-to-image translation problem. Through the use of a novel generator and discriminator architecture in conjunction with a composite loss function of L1, VGG19 based perceptual loss and adversarial loss, we were able to outperform the current state-of-the-art for image-to-image translation Pix2PixHD by more than 5dB PSNR. We validate this model by using it in the hologram optimisation pipeline, and displaying the results on our holographic display prototype. We show that our method can effectively compensate for optical aberrations and SLM errors, allowing us to outperform existing state-of-the-art methods.