Offline Arabic Handwritten Text Recognition for Unsegmented Words Using Convolutional Recurrent Neural Network

Abstract

This paper presents an analytical approach for offline Arabic Handwritten Text Recognition (HTR), based on Convolutional Recurrent Neural Network (CRNN). The suggested method is a three-part end-to-end trainable deep learning system that includes feature extraction, label prediction, and transcription part. The first part is performed by Convolutional Neural Network (CNN) layers, where sequential features are extracted. In the label prediction part, the extracted features are used to generate new sequential contextual features by feeding them to recurrent layers. This set of features for Arabic texts is then used to predict label distributions with fully connected layers. In the third part of the system, the transcription part, the predicted label distributions are translated into actual label sequences, using the Connectionist Temporal Classification (CTC) method. The experiments are carried out and reported on the publicly available IFN/ENIT database. The results of the proposed system are encouraging, and the recognition rates are comparable to those of numerous other systems in the literature.