Multidimensional Imaging: Towards Enhanced Spectrum, Polarization, and 3D Vision

Abstract

This thesis investigates three representative modalities of multidimensional imaging, including spectral imaging, polarization imaging, and 3D stereo imaging, which address the growing demand for efficient acquisition and processing of high-dimensional data in modern scientific and industrial applications. As the number of imaging dimensions increases, tra-ditional sequential scanning approaches face exponential growth in data volume and system complexity. In response, this work embraces a paradigm shift toward snapshot-based strategies that use optical encoding and computational decoding to enable compact, real-time, and high-fidelity multidimensional imaging. The first part of this thesis explores the reconstruction of hyperspectral images from RGB inputs. Starting from a lightweight convolutional neural network trained on synthetic data, we demonstrate accurate recovery of 31-band hypercubes from a single RGB image. To enhance reconstruction fidelity, a dual-RGB system using learned optical filters was developed,eliminating the need for sensors alignment. A dual-camera RGB-hyperspectral imaging system was built to collect a real-world dataset for training and evaluation. After that, we introduce a spectrally tunable light source and a scene-aware recovery framework, achieving improved results under controlled illumination conditions. These efforts collectively contribute to making hyperspectral imaging more affordable, flexible, and deployable. In the second chapter, we present advancements in color-polarization imaging using microgrid sensors. A demosaicking network was trained on real-world RGB-polarization image pairs captured using a dual-camera setup, addressing the limitations when using synthetic training data. Extending this work, we developed a polarization hyperspectral camera byintegrating a custom mosaic filter optimized via compressed sensing theory. The resulting system captures full hyperspectral data under multiple polarization states. Experimental validation using standard and polarimetric targets confirmed high reconstruction accuracy, while tests on biological samples demonstrated potential applications in biomedical polar-spectral analysis. The third chapter focuses on the development of a common viewpoint panoramic endoscope for 3D stereo colonoscopy. Designed as a 360-degree panoramic front-end attachment for standard colonoscopy, it features six optical subsystems arranged in a hexagonal configuration. A shared-viewpoint design was implemented using freeform optics and customZEMAX macros to simplify image stitching. High-precision diamond turning was used to fabricate concentric lens rings, and stray light suppression was achieved through targeted coating strategies identified via LightTools simulation. A real-time interface was developed for panoramic streaming, and a customized Structure-from-Motion algorithm enabled 3D surface reconstruction. This project demonstrates the challenges and rewards of integrating optical design, precision fabrication, and computational imaging into a robust, clinically relevant device. Together, these contributions provide practical solutions and theoretical insights into the design and implementation of snapshot-compatible, high-dimensional imaging systems, advancing the frontiers of low-cost, high-performance multidimensional sensing.