Inhaltszusammenfassung

Symmetries provide important cues for visual perception and object recognition in images. This paper explores the ability of product-unit neural networks to classify images with different symmetries and compares their performance with standard neural networks with the same number of artificial neurons. For this purpose two binary classification experiments were performed. In the first experiment, random patterns with either a horizontal or vertical symmetry axis were classified. In the second...Symmetries provide important cues for visual perception and object recognition in images. This paper explores the ability of product-unit neural networks to classify images with different symmetries and compares their performance with standard neural networks with the same number of artificial neurons. For this purpose two binary classification experiments were performed. In the first experiment, random patterns with either a horizontal or vertical symmetry axis were classified. In the second experiment, random patterns had to be distinguished from rotationally symmetric circular patterns. All networks achieved test accuracies of over 90 %, with the standard networks performing better in the axis symmetry experiment, but the product unit networks performing better in the rotation symmetry experiment. Possible implications of the results for object recognition and for solving symmetry classification problems are discussed.» weiterlesen» einklappen

Autoren

Klassifikation

DFG Fachgebiet:
Informatik

DDC Sachgruppe:
Technik