Residual Q-Learning Applied to Visual Attention

Abstract

Foveal vision features imagers with graded acuity coupled with context sensitive sensor gaze control, analogous to that prevalent throughout vertebrate vision. Foveal vision operates more efficiently than uniform acuity vision because resolution is treated as a dynamically allocatable resource, but requires a more refined visual attention mechanism. We demonstrate that reinforcement learning (RL) significantly improves the performance of foveal visual attention, and of the overall vision system, for the task of model based target recognition. A simulated foveal vision system is shown to classify targets with fewer fixations by learning strategies for the acquisition of visual information relevant to the task, and learning how to generalize these strategies in ambiguous and unexpected scenario conditions. 1 OVERVIEW OF FOVEAL VISION In contrast to the uniform acuity of conventional machine vision, virtually all advanced biological vision systems sample the scene in a space-variant fashion. Retinal acuity varies by several orders of magnitude within the