Visual Recognition using Neuromorphic Networks

Christoph Raschke
http://www.nd.edu/~crasche/
Postdoc in the Department of Psychology
University of Notre Dame

Abstract:

Visual object recognition occurs so rapid (<200ms), that one can assume that the initial object percept is a hypothesis. But what are useful descriptors for an object hypothesis? We propose that 2D spaces (regions) are an important component of object description. Such regions often represent surfaces of object parts, or silhouette features - the space between parts. We have developed a neuromorphic front-end that encodes space from gray-scale images using wave-propagating neuronal networks: A retina network signals contours by lines of spikes; subsequent cortical-like networks perform Blum's symmetric-axis transform, a region-encoding mechanism that transforms space into vectors useful for high-level representation. We discuss how the vectors maybe integrated to form object hypotheses.