Algorithms for Computational Vision and Sensing with Metasurface Lenses

Abstract

Vision systems transform light from the physical world into useful representations by combining carefully engineered optics with complementary algorithms. Their efficiency and accuracy depend critically on how well the optical and computational components are specialized to the task and co-designed as a unified system. Metasurfaces, a class of ultrathin optical elements composed of subwavelength-scale structures, have recently emerged as a versatile alternative to traditional refractive lenses. By encoding spatially varying phase or amplitude shifts at the nanoscale, metasurfaces allow compact lenses to be tailored for specific imaging tasks. In this dissertation, we explore the intersection of computer vision and applied physics by developing novel metasurface-based vision systems and learning-based methods to co-design their optical and algorithmic components. We demonstrate these methods on diverse tasks, including depth estimation, opto-electronic image processing, and snapshot hyperspectral imaging. Our designs are validated using both simulation and fabricated hardware prototypes. We present a unified, learning-driven framework for designing and modeling these systems, and show that complex visual tasks can be solved while introducing new trade-offs between optical complexity and computational effort.