Improving the Segmentation of Arabic Handwriting Using Ligature Detection Technique

Recognizing handwritten characters remains a critical and formidable challenge within the realm of computer vision. Although considerable strides have been made in enhancing English handwritten character recognition through various techniques, deciphering Arabic handwritten characters is particularly intricate. This complexity arises from the diverse array of writing styles among individuals, coupled with the various shapes that a single character can take when positioned differently within document images, rendering the task more perplexing. In this study, a novel segmentation method for Arabic handwritten scripts is suggested. This work aims to locate the local minima of the vertical and diagonal word image densities to precisely identify the segmentation points between the cursive letters. The proposed method starts with pre-processing the word image without affecting its main features, then calculates the directions pixel density of the word image by scanning it vertically and from angles 30 degrees to 90 degrees to count the pixel density from all directions and address the problem of overlapping letters, which is a commonly attitude in writing Arabic texts by many people. Local minima and thresholds are also determined to identify the ideal segmentation area. The proposed technique is tested on samples obtained from two datasets: A self-curated image dataset and the IFN/ENIT dataset. The results demonstrate that the proposed method achieves a significant improvement in the proportions of cursive segmentation of 92.96% on our dataset, as well as 89.37% on the IFN/ENIT dataset.