Task-Based Feature Learning and Enhancement for Bandwidth-Limited Applications

This research introduces an entirely new framework for vision processing that learns task-based visual features and enhances them in images to guide human action in vision-based tasks. We demonstrate a novel vision processing pipeline for prosthetic vision that learns navigation-based features in simulation via deep reinforcement learning and successfully enhances them in real world images. We further establish the efficacy of pre-trained, deep super resolution techniques in enhancing the resolution of satellite images despite a lack of domain-specific training, for earth observation and monitoring outcomes. In this thesis, we establish a new task-based vision processing paradigm for bandwidth-limited applications.