Representation of 2D objects with a topology preserving network

Abstract

We propose the use of a self-organizing neural network, the Growing Neural Gas, to represent bidimensional objects, due to its quality of topology preservation. As a result of an adaptative process, the object is represented by a Topology Preserving Graph, that constitutes an induced Delaunay Triangulation of their shapes. Features that are extracted from this graph simplify the later operations of classification and recognition, avoiding the high complexity of comparisons between graphs. This model of object characterization allows refining the quality of the representation based on the time available to its calculation, so that it will be the basis for the design of high performance real-time vision architectures. This work opens a new research field, because it employs the topology of a self-organizing neural network as feature, not, as usual, as a classifier.