Developmental Networks With Foveation

Abstract

The foveated nature of the human vision system (HVS) means the acuity on the retina peaks at the center of the fovea and gradually descends to the periphery with increasing eccentricity. Foveation is general-purpose, meaning the fovea is more often used than the periphery. Self-generated saccades dynamically project the fovea to different parts of the visual world so that the high-acuity fovea can process interested parts at different times. It is still unclear why biological vision uses foveation. This work is the first foveated neural network as far as we are aware, but it has a limited scope. We study two subjects here as follows. 1) We design a biological density of cones (BDOCs) foveation method for image warping to simulate a biologically plausible foveated retina using a commonly available uniform-pixel camera. 2) The subject of this article is not specific to tasks, but we choose a challenging task, visual navigation, as an example of quantitative and spatiotemporal tasks, and compare it with deep learning. Our experimental results showed that 1) the BDOC foveation is logically and visually correct; and 2) the developmental network (DN) performs better than deep learning in a surprising way and foveation helps both network types.