A Hybrid Recognition System for Islamic Annotation and Historical Arabic Handwritten Manuscripts

Abstract

In this paper, a multi neural-fuzzy recognition system with two combined statistical features, to solve the recognition problem of Historical Arabic Handwritten (HAH) manuscripts will be presented.  The first set of statistical features are center of mass, crosshair, outlier and blank ink histogram (CCOB).  The second feature is the principal component analysis (PCA). The new method will use two stages (levels) which are based on two classifiers, one public and one private, according to the similar features among characters. In the first level, we built a public classifier to deal with all character groups. Each group contains characters with overlapped features. The public classifier classifies the characters in the segmented character data set, which is captured from HAH manuscripts to specified groups. In the first stage, the system was applied to 34 Arabic characters and achieved 97.15% recognition rate for the tested dataset. In the second level, we created a private classifier for each group to recognize and classify the characters within a group which achieved 99.34% recognition rate for the tested dataset using the two level model.