Neural network architectures for rotated character recognition

In this paper we explore a neural network (NN) approach that is analogous to the human's straightforward pattern matching, where some rotation is taking place in high level neurons close to symbols. The main objective here is to develop ideas to simulate the rotation and verify them by using a large number of handwritten characters. We propose a feed forward NN where the links between input and hidden units are locally connected and weights are symmetrically shared. In the recognition process the total input values (except biases) to hidden units are rotated according to the number of possible orientations and the activation values of output units are calculated for each orientation to find the best output. Since the iteration is done only at high level NN, the computation is not time consuming. We conducted experiments using 6400 handwritten numeric characters and obtained a recognition rate of 90.1 % for eight orientations.