Online Persian/Arabic Writer Identification using Gated Recurrent Unit Neural Networks

Abstract

Conventional methods in writer identification mostly rely on hand-crafted features to represent the characteristics of different handwritten scripts. In this paper, we propose an end-to-end model for online text-independent writer identification on Persian/Arabic online handwritten scripts by using Gated Recurrent Unit (GRU) neural networks. The method does not require any specific knowledge for handwriting data analysis. Because of the exclusive ability of deep neural networks, we just represented our data by Random Strokes (RS) representations, which are differential horizontal and vertical coordinates extracted from different handwritings with a predefined length. This representation is a context independent representation. Therefore, this writer identification at RS level is more general than character level or word level in identification systems, which require character or word segmentation. The RS representation is then fed to a GRU neural network to represent the sequence for final classification. All RS features of a writer are then classified independently, and in the final stage, the posterior probabilities are averaged to make the final decision. Experiments on KHATT database, which consists of online handwritings of Arabic writers, gave us 100% accuracy on 10 writers and 76% accuracy on 50 writers, which is much better than previous works on online Persian/Arabic writer identification.