Evolving a Neural Network Active Vision System for Shape Discrimination

Previous research has demonstrated the potential for neural network controlled active vision systems to solve shape discrimination and object recognition tasks. However, this approach has not been very well explored, and previous implementations of such systems have been somewhat limited in scope. We present an evolved neural network based active vision system that is able to move about a 2D surface in any direction, along with the ability to zoom and rotate. We demonstrate that a system with such features can correctly classify shapes presented to it, despite variance in location, scale, and rotation. And, contrary to our initial assumptions, effective discrimination is actually improved when the ability to rotate is disabled.