Abstract
This paper presents a rotation invariant neural pattern recognition system capable of recognizing a rotated pattern and estimating its rotation angle. The system consists of two networks, the first and the second networks. The first one is to recognize a rotated pattern and to estimate its rotation angle. Next, the second one is to estimate the rotation angle more exactly. Those networks detect in their preprocessors edge features which are input patterns to multilayered networks. They are not insensitive to rotation. Rotational invariance is achieved in the multilayered networks, which are the three-layered networks. In the first network an output layer is divided into two different parts. One is to recognize an input pattern to be invariant to rotation and the other is to estimate its rotation angle. The second network can compute smaller rotation angle by using the popular three-layered network.
In the paper, it is shown that, by means of computer simulations on a binary pattern recognition problem, the proposed system is able to recognize rotated patterns, and further can estimate their rotation angle.
