Type a line, pick how the page arrived, then press Read.
Tesseract 3 (the classic engine, still inside today's Tesseract as --oem 0) cuts every word into blobs of ink, classifies each piece by its shape features, and repairs mistakes with a dictionary. Tesseract 4 and 5 (--oem 1) read the whole line with the LSTM. The page layout step is the same for both.
Run both engines, then click a word on the classic page.
Stock Tesseract was trained on clean book and web text. Docupath's documents arrive as scans, faxes and PDFs full of ruled tables, in Docupath's own fonts. We took Tesseract's own network out of eng.traineddata, trained it further on Docupath-style lines, and wrote the new weights back into the same file format. The real Tesseract engine on the Page reader tab loads that file.
Docupath Cookie Policy (website-facing policy, from Sprinto). Only this public-facing policy is used here. Sections 7 to 10 and every third table row were held back for testing and never trained on.
A rough fax of Cookie Policy text from the held-out sections (never trained on), read end to end by the real Tesseract engine, once with each model. In the readings, red is wrong, and green marks a letter stock Tesseract got wrong that the fine-tuned model got right. Orange is a letter stock got right and fine-tuning broke. Lines with a green edge improved.
Each datapoint is one line image and its exact text. Lines are cut from the policy and rendered the way Docupath documents really arrive. Click one to watch both networks read it.