Maxout into MDLSTM for Offline Arabic Handwriting Recognition

Research on Arabic handwriting recognition has been seriously challenged due to its cursive appearance, the variety of writers and the diversity of styles. In fact, motivated by a series of success cases in computer vision, we try to explore the Maxout units in Multidirectional neural networks for the offline task. Therefore, in this work, we model an Arabic handwritten word with a deep MDLSTM-based system. This architecture can directly work on raw input images since it enables us to model the script variations on both axes of the image due to recurrence over them. However, several problems, such as the vanishing gradient, can affect the training of this recognition system. To overcome this problem, we should integrate Maxout units into MDLSTM system in order to enhance it and improve its performance. In this context, different integration modes are carried out to draw out the best topology. Proposed systems are evaluated on a large database IFN/ENIT. According to the experimental results and compared to the baseline system, the best tested architecture reduced the label error rate by 6.86%.