UC San Diego

Computer vision has seen great change in the last decade, characterized by shallow machine learning models on hand-tuned features in 2010, to deep models with end-to-end learned features in 2018.

This dissertation includes work from both the pre- and post-deep-learning periods. It includes two chapters on local descriptors - functions of local regions of images, designed to produce objects with desirable properties when compared with particular distance functions.

The first of these chapters describes the LUCID descriptor, a descriptor designed to run quickly on devices with limited floating-point support (e.g. mobile phones circa 2012). The second describes match-time covariance, a framework for building descriptors, which delegates to the distance computation the task of ensuring rotation and scale invariance. It also describes an instance of such a descriptor. Both these descriptors were hand designed, leveraging expertise in image formation and algorithms to create functions which would extract desired image information while discarding nuisance information.

The final project presented herein is a deep learning system for computational microscopy. It uses the paradigm of image-to-image translation to predict hard-to-acquire fluorescence images from easy-to-acquire transmitted light images, enabling biological assays which were previously impossible. In this domain, accuracy is paramount while speed is secondary, necessitating the development of a new deep learning model for the domain of image-to-image translation.