Fuzzy Hough transform and an MLP with fuzzy input/output for character recognition

A neuro-fuzzy system for character recognition using a fuzzy Hough transform technique is presented in this paper. For each character pattern, membership values are determined for a number of fuzzy sets defined on the standard Hough transform accumulator cells. These basic fuzzy sets are combined by t-norms to synthesize additional fuzzy sets whose heights form an ndimensional feature vector for the pattern. A 3n-dimensional fuzzy linguistic vector is generated from the n-dimensional feature vector by defining three linguistic fuzzy sets, namely, weak, moderate and strong. The linguistic set membership functions are derived from the Butterworth polynomials and are similar to the gain functions of low pass, band pass and high pass filters, respectively. A multilayer perceptron (MLP) is trained with the fuzzy linguistic vectors by the back propagation of errors. The MLP outputs represent fuzzy sets denoting similarity of an input feature vector to a number of character pattern classes. Recognition accuracy of the system is more than 98%.