Nonlinear Normalization of Input Patterns to Handwritten Character Variability in Handwriting Recognition Neural Network

The issue of input variability resulting from writer changes is one of the most crucial factors influencing the effectiveness of handwritten character recognition systems. A solution to this problem is adaptation or normalization of the input, in a way that all the parameters of the input representation are adapted to that of a single writer, and a kind of normalization is applied to the input pattern against the writer changes, before recognition. This paper propose such a method that uses a feed forward nonlinear auto associative Neural network that is trained for mapping character pictures to a normal set of pictures as the desirable output. Then all reconstructed pictures are given to a feed forward neural network classifier in order to recognize each of the character´s class. In the second method with inspiration from processing in human brain, we add a reverse network to adaptation network [Cortex]. Given an input our forward model generates an initial hypothesis (bottom-up processing). This model extract the context of current picture in middle layer, then the reverse network receive this context and process it (top-down processing). Output of the mentioned reverse network is entered to the decoding layer of forward network and influence the output. By adding the inverse network to recognition model, it is seen that recognition rate is reached to 99.55% on test data set of IFHCDB [1] that have improvement in comparison with the recent works.