Codebook Based Handwritten and Printed Arabic Text Zone Classification

In this work, we present a method for classifying handwritten and printed Arabic text zones in noisy document images. We use Three-Adjacent-Segment (TAS) [8] based features which capture properties of a script. We construct two different codebooks of the local shape features extracted from a set of handwritten and printed Arabic documents and use it to train both Support Vector Machine and Fisher’s linear discriminant classifiers using normalized histograms. Due to robustness of TAS features to noise the proposed classification scheme is suitable for noisy document images where performance of other methods degrades drastically. Our experiments show that we can achieve 90–95% classification accuracy. This method is also robust to segmentation results which may contain segments at word, line or paragraph level.