Today we are releasing a major update to OpenDrawing's text recognition engine. The new model delivers significantly improved accuracy on handwritten annotations and legacy typed schematics, reducing manual correction time by 80% on archives that previously required the most review effort.
What Changed
The previous OCR model performed well on clean, typed annotations from modern CAD-exported drawings but degraded on two common legacy drawing types: drawings with hand-lettered callouts (prevalent in drawings from the 1960s through 1980s) and drawings produced on older typewriters or dot-matrix plotters with inconsistent character spacing.
The new model was trained on a significantly expanded dataset of legacy engineering drawings spanning five decades of drafting styles, including hand-lettered relay schematics, typewritten equipment schedules, and rubber-stamp annotations common in utility and industrial archives.
What This Means in Practice
For customers with legacy paper archives, the practical impact is fewer items in the manual review queue. A drawing package that previously produced 40 low-confidence text annotations requiring human confirmation now produces 8 to 10. Review time per drawing drops from 25 to 30 minutes to under 10 minutes for most legacy drawing types.
The improvement is most pronounced on three annotation types: device number tags (e.g., "87T-1A"), handwritten equipment specifications (transformer kVA ratings, cable sizes), and rubber-stamped title block data.
Who This Affects
All existing customers automatically have access to the updated model, no configuration changes are required. New drawings submitted from today will be processed with the new engine. For customers who want to reprocess previously-digitized drawings, the OpenDrawing team can arrange batch reprocessing on request.