Rotated-Pattern Normalization by Neural Network

Abstract

This paper discusses a neural network for a rotation invariant pattern recognition system. The proposed network can normalize a rotated random pattern into the unrotated standard pattern. The normalization is achieved by cascading a preceding rotation-angle extracting network and the succeedingnormalizing network. This basic structure is the same as the previously reported shifted-pattern normalizing network. The network is trained to normalize input random pattern direction into its center of gravity downwardsor in some other predetermined direction. The weight distribution viewed from the input layer in each hidden unit reveals a Fourier transform like in the radius and angle directions.