Optical Character Recognition of Historical Moroccan ArabicCalligraphy Using Transformer-Based TrOCR

Historical Moroccan Arabic manuscripts represent an importantcultural heritage; however, their automatic transcription remainsa challenging task due to document degradation, complex calligraphic styles, touching and overlapping characters, and largevariations in handwriting. Traditional Optical Character Recognition (OCR) methods and conventional deep learning approachesoften rely on accurate character or word segmentation. However,in historical Moroccan Arabic calligraphy, character segmentation is particularly difficult because adjacent characters are frequently connected or overlapped, making character-level recognition unreliable.To address this challenge, this paper proposes a Transformerbased OCR framework for historical Moroccan Arabic calligraphy. First, historical manuscript pages are preprocessed to improve image quality and automatically segmented into individualtext lines using a dedicated preprocessing pipeline. Instead ofperforming explicit character segmentation, each text line is directly recognized using a fine-tuned TrOCR model based on theTransformer architecture. Using the self-attention mechanism,the proposed approach effectively models long-range contextualdependencies across the entire text line, eliminating the need forcharacter-level segmentation.The proposed model is trained and evaluated on a historicalMoroccan Arabic calligraphy data set using the Character Error Rate (CER) as the evaluation metric. Experimental resultsdemonstrate that the Transformer-based approach provides an effective solution for recognizing historical Moroccan manuscriptsand highlight the advantages of Transformer architectures overconventional deep learning methods for complex handwrittendocument recognition. This work contributes to the digital preservation of the Moroccan documentary heritage by providing an accurate OCR framework for historical Arabic manuscripts.