Fuzzy modeling for handwritten Arabic numeral recognition

In this paper we present a novel fuzzy technique for Arabic (Indian) online digits recognition. We use directional features to automatically build generic fuzzy models for Arabic online digits using the training data. The fuzzy models include the samples’ trend lines, the upper and lower envelops of the samples of each digit. Automatically generated weights for the different segments of the digits’ models are also used. In addition, the fuzzy intervals are automatically estimated using the training data. The fuzzy models produce robust models that can handle the variability in the handwriting styles. The classification phase consists of two cascaded stages, in the first stage the system classifies digits into zero/nonzero classes using five features (viz. length, width, height, height’s variance and aspect ratio) and the second stage classifies digits 1 to 9 using fuzzy classification based on directional and segment histogram features. Support Vector Machine (SVM) is used in the first stage and syntactic fuzzy classifier in the second stage. A database containing 32695 Arabic online digits is used in the experimentation. The results show that the first stage (zero/nonzero) achieved accuracy of 99.55% and the second stage (digits from 1 to 9) achieved accuracy of 98.01%. The misclassified samples are evaluated subjectively and results indicate that humans could not classify = 35% of the misclassified digits.