Symmetry feature extraction for artificial neural networks: providing position, size, rotation invariance

Pattern recognition systems, such as those implemented using Artificial Neural Networks, have not yet reached the performance of a human. Better mechanisms are required to identify a pattern regardless of any size, rotation or position variation. Symmetry is a property of many of the objects in the environment. The main research problem will be to investigate the effectiveness of symmetry feature based pattern classification methodologies. Based on this goal, a set of hypotheses of this research will be investigated: 1) Symmetry features of patterns can be used for classification; 2) A symmetry feature based mechanism for representing patterns is invariant under certain transformations; 3) Classification of a pattern's symmetry features can improve generalization. The research will develop symmetry feature extraction methodologies along with the associated mathematical and algorithmic techniques for enhancing the generalization ability of artificial neural networks.