An Evolutionary Active-Vision System

We describe an evolutionary vision system capable of autonomously scanning through an image while zooming in and out and changing filtering strategy in order to perform shape discrimination. The system consists of a small artificial retina controlled by an evolutionary recurrent neural network without hidden units. We show that such a simple active-vision system can successfully recognize different shapes independently of their position and size by dynamically exploring relevant parts of the image. We also show that a standard feed-forward neural network trained with the back-propagation algorithm cannot perform the task, not even with hidden units added to the architecture. Given its compactness, computational requirements, and versatility, this evolutionary active vision system is a suitable solution for small-size and embedded vision systems with stringent energetic and computational requirements, such as micro-robotic systems. In addition, this approach provides a framework for studying emergent active-vision behaviors in autonomous systems.