Local features enhancement using deep auto-encoder scheme for the recognition of the proposed handwritten Arabic-Maghrebi characters database

Over the years, automated handwritten Arabic character recognition systems have evolved. However, optical Arabic character recognition systems still suffer from low performances in the wild because of high human handwriting variations, styles, ambiguity and complexity. One of the main contributions of this paper is to create and present in detail a new database for Handwritten Arabic-Maghrebi Characters (HAMCDB), this handwriting style, which is represented and used for the first time in this field of character recognition and is much more difficult, poses additional challenges and complexities due to its characteristics. The database consists of all shapes of Arabic characters written in the Maghrebi style. The samples in the database were obtained from a variety of sources, the most important of which was the Algerian manuscripts portal (http://pam.univ-adrar.edu.dz/), which is a platform designed by the work team of the Algerian Manuscript's laboratory in Africa, to safeguard the humanitarian patrimony, where the reader can view and download digital copies and through this work by offering an OCR that is specific to the style of our region, we hope to make the search's operation more easier. HAMCDB understands a total of 1560 character images with 78 shapes and 20 images for each one. We propose a new and efficient Arabic handwritten character recognition scheme, tested on two datasets; the new proposed dataset HAMCDB compared to the public database AHCD of handwritten Arabic characters, where the local features using Gabor filter, Histogram of Oriented Gradients (HOG) and Binarized Statistical Image Features (BSIF) are enhanced with a deep auto-encoder architecture for better accuracy. These features are finally fed to a classification process based on nearest neighbor's classifier and cosine Mahalanobis distance to obtain final character labels. Experimental analyses on HAMCDB and AHCD databases, versus existing methods demonstrate the efficacy of the presented framework, and our framework's performance is promising, achieving accuracies of 98.63% and 98.95% using Gabor filters for HAMCDB and AHCD respectively, and 100% with HOG and BSIF.

Automated summary

This paper introduces a new database HAMCDB for handwritten Arabic-Maghrebi characters and a efficient recognition scheme. With the use of local features and deep auto-encoder, the scheme achieves high accuracy on different datasets.