An Arabic handwriting synthesis system Yousef Elarian a , Irfan Ahmad a , Sameh Awaida b , Wasfi G. Al-Khatib a , Abdelmalek Zidouri a,n a King Fahd University of Petroleum and Minerals, Dhahran 31261, Saudi Arabia b Qassim University, Qassim 51452, Saudi Arabia article info Article history: Received 3 May 2014 Received in revised form 7 August 2014 Accepted 15 September 2014 Keywords: Arabic handwriting Handwriting synthesis Training data Offline text recognition Statistical model abstract In this paper, we present an Arabic handwriting synthesis system. Two concatenation models to synthesize Arabic words from segmented characters are adopted: Extended-Glyphs connection and Synthetic-Extensions connection. We use our system to synthesize handwriting from a collected dataset and inject it into an expanded dataset. We experiment by training a state-of-the-art Arabic handwriting recognition system on the collected dataset, as well as on the expanded dataset, and test it on the IFN/ ENIT Arabic benchmark dataset. We show significant improvement in recognition performance due to the data that was synthesized by our system. & 2014 Elsevier Ltd. All rights reserved. 1. Introduction Handwriting recognition is an active area where researchers are trying various approaches to increase recognition rates [1]. Researchers agree that expanding the training set of a text recognition system is generally beneficial to recognition rates. However, conventional ways of collecting datasets can be time- consuming and may incur a lot of effort, especially for ground- truthing. Hence, researchers proposed the use of synthesized data in expanding training sets of recognition systems [2–6]. Handwriting synthesis refers to the computer generation of online and offline data that resemble human handwriting. It is a reverse process for handwriting recognition as it transforms input text into image samples, whereas recognition maps handwritten samples into digital text. Handwriting synthesis has become a topic of rapidly increasing interest because of its applications such as the improvement of text recognition systems (in terms of overall performance [2,7], stability [8,9], and speed [10,11]), personalized fonts [12,13], and forgery detection [14,15]. Depending on the application, synthesis methods and their corresponding evaluation methods vary. Perso- nalized fonts, for example, aim at capturing the style of a particular writer and tend to be evaluated subjectively. Whereas synthesized data for text recognition may aim at maximizing style variability within natural limits [3,16,17] and its evaluation is mainly tied to recognition rates. Handwriting synthesis encompasses generation and concatena- tion operations [18,19]. Handwriting generation operations alter samples of handwriting to increase their shape-variability within some closed-vocabulary. Concatenation operations, in contrast, aim at the compilation of new units of vocabulary, such as words, from a smaller pool of basic samples, such as characters. Handwriting generation can be seen as the inverse operation of preprocessing in a text recognition system whereas handwriting concatenation can be regarded as the inverse operation of segmentation. Synthesized data can improve systems that have deficiencies in their text segmentation accuracy, their recognition features and classifiers, or in the variability of their training data. One advan- tage of this approach is that it functions on the data level which is system-independent. Arabic is a widely used language and the Arabic script is used in other languages as well like Urdu and Persian [20]. In Arabic, most characters must connect to their successor within a word. These characters take one of four character-shapes: Beginning (B), Middle (M), Ending (E), and Alone (A); the few characters that do not connect to their successors can only take the (E) or (A) character-shapes. These characters cause Arabic words to break into Pieces of Arabic Words (PAWs). From right to left, a multi- character PAW consists of one (B) character-shape followed by zero or more (M) character-shapes and is terminated by one (E) character-shape. A PAW that consists solely of one character always takes the (A) character-shape. Characters connect in Arabic via a stroke called the Kashida [21]. Kashida are semi-horizontal strokes that often lie in the Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/pr Pattern Recognition http://dx.doi.org/10.1016/j.patcog.2014.09.013 0031-3203/& 2014 Elsevier Ltd. All rights reserved. n Corresponding author. Fax: þ966 138603535. E-mail addresses: yarian@kfupm.edu.sa (Y. Elarian), irfanics@kfupm.edu.sa (I. Ahmad), s.awaida@qu.edu.sa (S. Awaida), wasfi@kfupm.edu.sa (W.G. Al-Khatib), malek@kfupm.edu.sa (A. Zidouri). Please cite this article as: Y. Elarian, et al., An Arabic handwriting synthesis system, Pattern Recognition (2014), http://dx.doi.org/ 10.1016/j.patcog.2014.09.013i Pattern Recognition ∎ (∎∎∎∎) ∎∎∎–∎∎∎