Recognition of rotated patterns using a neocognitron
Soichirou Satoh, Jousuke Kuroiwa, H. Aso, Sei Miyake · 1999
A rotation-invariant neocognitron is constructed by extending the neocognitron which can recognize translated, scaled and/or distorted patterns from training ones. In constructing the model two technical methods during the learning, a "threshold-controlling method" and a "rotation matrix method", are proposed. In numerical simulations, it is shown that the model can recognize globally and/or locally rotated patterns in an arbitrary angle without learning the patterns themselves. 1 Introduction It is very important to realize recognition systems insensitive to different kinds of scaling, translation, distortion and rotation. Many neural networks have been proposed to achieve this purpose. Neocognitron[Fukushima, 1988] is a multi-layered neural network model for pattern recognition which is considerably robust against distortion, scaling, and/or translation of patterns. After unsupervised learning, it can recognize input patterns without being affected by distortion, change in these siz...