Using a Secondary Network for Foveation in Computer Vision

Abstract

Foveation plays a important role in the human visual system, but is currently hardly used in computer vision. Even though the "input" and systems used in computer vision are quite di erent from human vision, it is plausible foveation provides some bene ts. In order to investigate this I compared classi cation losses of foveated images and non-foveated images. In addition I trained a Deep-Q Network to learn which foveation locations are optimal for a given image. Results show that there is a bene t of foveation, and that it is possible for a computer vision model to learn which foveation locations are optimal.