An Innovative Ground Truth Dataset for Automated Validation of Arabic Handwritten Character Segmentation Algorithms

Abstract

Character segmentation is one of the most critical phases of Arabic handwriting recognition systems. The validation of Arabic Handwritten Character Segmentation (AHCS) algorithms requires Ground Truthed Datasets (GTD) that provide information on how characters should be segmented. Existing Arabic handwritten datasets provide Ground Truth Files (GTFs) which describe many entries, however, up to date, there is a lack of character-level information. Very few attempts have been made in the literature to establish character-level GTFs. Existing attempts provide characters’ boundaries or unique Segmentation Points (SP). The concept of constant SPs does not go in line with Arabic script features which suppose the presence of many possible SPs between two successive characters. Consequently, existing GTDs cannot allow reliable validation for all varieties of AHCS algorithms. In this paper, we propose a new ground-truthing concept that captures Segmentation Areas (SA) instead of SPs. This concept could respond to all varieties of AHCS algorithms and provides information about overlapping/touching characters and vertical ligatures. Our proposed GTD consists of 1400 GTF describing character-level information for a subset of 1400 images from the IFN/ENIT database. The proposed dataset is publicly available at (https://www.kaggle.com/datasets/mohsineelkhayati/character-level-ground-truth-dataset-for-ifnenit)