A Comprehensive Evaluation of YOLO Networks for Extracting Words from Persian Handwritten Text

Optical Character Recognition in the case of handwritten text is a critical challenge in web-based applications, such as form processing, digital archiving, and elearning platforms. This study addresses the initial phase of this process-word detection within Persian handwritten texts-using advanced YOLO architectures. The Sadri dataset was employed, comprising 500 pages of handwritten Persian text written by 250 men and 250 women. From this, 102 representative pages (51 men, 51 women) were randomly selected and annotated with bounding boxes for each word, preserving handwriting diversity. Multiple YOLO models were trained and evaluated, with YOLOv8x achieved the highest mAP@ 50 of 97.40 % and mAP@ <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$50-95$</tex> of 74.20 %. Additionally, YOLOv9m, with only 20.1 million parameters (less than onethird of YOLOv8x's 68.2 million), achieved efficient mAP@ 50 of 96.40% and mAP@ <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$50-95$</tex> of 72.30 %. These results demonstrate the potential of YOLO models for accurate and efficient detection of Persian handwritten words, enabling realtime and scalable solutions for diverse web applications.