قراءة أربعين عاماً من التقارير الحقلية الفارسية

A typical national operator holds tens of terabytes of scanned documents, most of them in Persian or Arabic, spanning four decades of tools and scan quality. General-purpose OCR engines handle clean typed Persian adequately and degrade sharply everywhere else.

The pipeline that works is staged: deskew and denoise, then layout analysis, then recognition with separate models for printed Latin, printed Persian and Arabic script, tables, and handwriting. Treating the page as one recognition problem is where quality is lost.

The uncomfortable finding is consistent across archives. Handwritten remarks columns, which are where crews recorded what actually happened, are recognised at materially lower accuracy than the typed body of the same page. Those regions are flagged and routed to human verification rather than accepted.

Confidence scoring per region is therefore not a nicety. Without it, a structured extraction built on the OCR output inherits errors it cannot see, and a database of confidently wrong numbers is worse than no database.

The realistic outcome of an archive programme is a searchable corpus with a known and reported error profile, plus a verification queue that a records team works through over months. That is a large improvement over a cupboard, and it is not the same as perfect digitisation.