Visual Attention for Foveal Machine Vision

An attention mechanism can influence system behavior and performance only to the extent in which system resources are allocatable. A key visual resource is acuity, which is fixed in conventional uniform acuity machine vision. Foveal vision features imagers with graded acuity coupled with context sensitive sensor gaze control, analogous to that prevalent throughout vertebrate vision. Foveal systems operate more efficiently than uniform acuity systems because resolution is treated as a dynamically allocatable resource. Wide FOV and localized high acuity are simultaneously supported while minimizing sensor data to that which is relevant. Foveal system development is hampered by the need for attention mechanisms and gaze control which are more refined than that of uniform acuity vision, and must rapidly allocate spatiotemporal resolution in dynamic non-deterministic environments. This program will apply reinforcement learning to foveal attention and gaze control to improve the performance of autonomous platform vision. Reinforcement learning will be used to optimize sequences of salient point interrogations in scale space e.g., saccades), and improve the information acquisition ability of a foveal agent. Visual attention will be posed as a optimization problem which attempts to accelerate foveal object recognition and reduce multiresolution object model complexity. Reinforcement learning will also be used to control the switching of vision task modalities, and preattentive preemption. The learning of attention invariants among different targets and poses will also be investigated.