Improving the DBLSTM for on-line Arabic handwriting recognition

Abstract

Various applications involved in the computer recognition of pen-input handwritten words, such as the online form filling, text editing, note taking, and so on. Therefore, a great deal of research work tries to improve the recognition rate of those online words recognition systems resulting in several effective methods. Relevant results related to Latin and Chinese scripts have been achieved. However, the Arabic script has been neglected so far, which stimulated us to propose a new online Arabic handwriting recognition system based on DBLSTM that relies on three techniques that were applied in order to enhance its performance. First, the dropout was applied, in different positions, to prevent overfitting. Then, ReLU and Maxout units were added, in different ways, to overcome the vanishing gradient problem. These proposed systems were tested on a large database ADAB to prove its performance against difficult conditions such as the variety of writers, the large vocabulary and the diversity of style. According to the experimental results and compared to the baseline system, the best tested architecture gives a reduction of 10.99% in label error rate.