Recurrent Neural Network Method in Arabic Words Recognition System

TL;DR

This paper presents a recurrent neural network approach with connectionist temporal classification for online Arabic handwriting recognition, achieving higher accuracy than previous methods.

Contribution

Introduces a recurrent neural network with CTC for Arabic word recognition, improving accuracy over traditional HMM-based systems.

Findings

Recognition rate of 79% for Arabic words

Outperforms previous HMM-based system with 70% accuracy

Demonstrates effectiveness of RNN with CTC in complex handwriting recognition

Abstract

The recognition of unconstrained handwriting continues to be a difficult task for computers despite active research for several decades. This is because handwritten text offers great challenges such as character and word segmentation, character recognition, variation between handwriting styles, different character size and no font constraints as well as the background clarity. In this paper primarily discussed Online Handwriting Recognition methods for Arabic words which being often used among then across the Middle East and North Africa people. Because of the characteristic of the whole body of the Arabic words, namely connectivity between the characters, thereby the segmentation of An Arabic word is very difficult. We introduced a recurrent neural network to online handwriting Arabic word recognition. The key innovation is a recently produce recurrent neural networks objective function known as connectionist temporal classification. The system consists of an advanced recurrent neural network with an output layer designed for sequence labeling, partially combined with a probabilistic language model. Experimental results show that unconstrained Arabic words achieve recognition rates about 79%, which is significantly higher than the about 70% using a previously developed hidden markov model based recognition system.