Neural Dispersive Hologram for Computational Submillimeter-wave Imaging

Abstract

Conventional submillimeter-wave imaging systems struggle to achieve scalability, real-time performance, and low complexity simultaneously. Microwave computational imaging offers a solution by using frequency diversity to image scenes through various measurement modes, shifting complexity from hardware to software. Typically, these systems first design the element that creates diverse illumination and then tune reconstruction algorithm parameters for good image quality.

In contrast to the sequential design approach, this thesis proposes a joint optimization method for a frequency-diverse phase hologram and the reconstruction neural network parameters in submillimeter-wave imaging. The goal is to optimize the hologram pattern for enhanced frequency diversity and its corresponding neural network for improved image quality.

First, a literature review of computational imaging systems from optics to millimeter-wave is presented. Additionally, differentiable imaging theory and recent work in the joint optimization of computational imaging systems are studied. A PyTorch-based Fourier-optics simulation codebase is developed to model the imaging physics process, mapping the reflectivity of objects to the reflected frequency response. This codebase analyzes the impact of joint optimization on efficiency, frequency diversity, and reconstruction quality, resulting in two optimal holograms with minimal unit sizes of 1 mm and 2 mm for manufacturing.

To validate the performance of the proposed computational imaging system with holograms, a quasi-optics setup operating at 220-330 GHz is built. While few similarities were observed between simulated and measured field patterns, scanning the fields near the actual hologram and the corresponding comparisons suggest that the designed holograms work as expected. The measured frequency diversity of the designed holograms exceeded that of both the simulations and the previous design. The comparison of measured and simulated frequency responses, along with imaging experiments, indicates partial validity of the current imaging physics model. Future work should focus on developing a more accurate model that accounts for non-ideal factors to better match the measured and simulated frequency responses.