Projectivity invariant pattern recognition with high-order neural networks

Abstract

The need to provide a neural network for pattern recognition with invariance to a new transformation, projectivity, is considered. This invariance is justified when working with object images that can appear rotated in relation to an axis contained in its own plane. An invariable relation to the transformation is found, the double ratio of four points, and incorporated to the network as a restriction to the weights. A projectivity invariant pattern classifier has been simulated. Besides, some considerations about high order neural networks are expounded.