A Comparative Study on Optical Modeling Units for Off-Line Arabic Text Recognition

The role of the optical model in a text recognition system is to model the textual information written in image documents. This paper compares the performance of four Arabic optical modeling units in a Multi-Dimensional Long Short-Term Memory based state-of-the-art Arabic text recognition system. These units are: 1) The isolated characters, 2) Extended isolated characters with the different shapes of Lam-Alef (), 3) The character shapes within their contexts and, 4) The recently proposed sub-character units that allow sharing similar patterns in the different character shapes. Experiments are conducted on six tasks using Maurdor and Khatt databases. For a fair comparison, optical models are trained from scratch. The decoding is performed using 1) the predictions of the optical model only and, 2) combined with a 3-gram hybrid word/part-of-Arabic word language model. Results in terms of Word Error Rate show that best results are generally obtained with systems using isolated characters as the basic modeling units, although differences in the performance among different systems are negligible.