Data-Efficient Machine Learning for Computational Imaging

Abstract

This thesis presents a method that improves data efficiency in computational imaging by incorporating prior knowledge from physical models into machine learning algorithms. Our approach optimizes image reconstruction from sparse and noisy datasets by utilizing physical constraints to guide deep learning models. This integration accelerates the imaging workflow, minimizes the need for large datasets, and improves resilience to measurement noise. The key insight is that physical model-based priors can regularize deep learning for more robust performance. Experiments demonstrate how this physics-assisted machine learning technique enables faster, more accurate, and reliable imaging. By facilitating high-quality imaging from limited data, this method has the potential to advance applications in healthcare, material studies, and industrial inspection. One of the highlights of our method is the application of real-time 2D imaging for improving 3D printing. High-performance manufacturing is achieved by training a neural model combined with a system of dynamic equations. The thesis offers a framework that seamlessly integrates physical insights and data-driven methods, enabling advances beyond what either approach could achieve alone.