Off-line Handwritten Arabic Text Recognition using Convolutional DL Networks

Abstract

In recent decades many researchers worked on handwritten document analysis field and more specifically for Arabic Handwriting Script (AHS). Deep learning (DL) has revolutionized computer vision with several good examples, particularly the studies of the convolutional neural network on image classification. In this paper we investigate the benefit of deep convolution networks in textual image classification. Convolutional Deep Belief Networks (CDBN) is applied to learn automatically the finest discriminative features from textual image data consisting of AHS. This architecture is able to lay hold of the advantages of Deep Belief Network and Convolutional Neural Network. We subjoin Regularization methods to our CDBN model so that we can address the issue of over-fitting. We evaluated our proposed model on low and high-level dimension in Arabic textual (character /word) images using IFN/ENIT datasets with data augmentation. Experimental results show that our proposed CDBN architectures achieve better performance.