Arabic Manuscripts Alignment, Segmentation, Recognition, and Classification

This research proposes a novel approach for aligning, segmenting, recognizing, and classifying Arabic manuscripts. The alignment process involves removing noise, correcting rotation and skewing, and aligning the manuscripts with a standard template. This alignment is crucial in determining whether a manuscript is known or unknown. The proposed approach, called Automated Arabic Manuscript Alignment (ArM), takes multiple input images, aligns them, refines their internal parameters, and analyzes each parameter. It then recognizes and maps the regions of interest (RoI) in the manuscript to the corresponding printed or handwritten content. Finally, a suitable optical character recognition (OCR) system is used to recognize the content within each region. The first phase of this research focuses on manuscript alignment to OCR historical and ancient manuscripts. In the second phase, the alignment is combined with OCRing to extract and recognize the text from each RoI in the Arabic manuscripts. Additionally, the classification of Arabic manuscripts into known or unknown categories is accomplished through a digital image processing technique, specifically an analysis of Arabic cards and the extraction of RoI. The aim is to make the entire process more efficient, accurate, and faster. Several experiments were conducted on different regions of the manuscript using the intersection over union (IoU) metric. The default Arabic OCRing achieved the best results with an accuracy of 99.69%. To evaluate the performance of this technology compared to the state-of-the-art, it relies on multiple interconnected modules within the OCRing processing pipeline, including image preprocessing, layout analysis, line and word segmentation, and character recognition. This paper surveys the evaluation process for Arabic OCRing, highlighting the importance of available tools and their influence on the indicator metrics used. The specification and preparation of the benchmarking dataset are discussed, as it significantly impacts the evaluation results. Overall, the comprehensive evaluation of OCRing, accounting for the various subsystems, and the careful consideration of the benchmarking dataset are essential for accurate performance assessment compared to the current progress in the field.