NetroDoc test notes

OCR Accuracy: What Our Controlled Scan Tests Actually Show

OCR accuracy depends on the input. Instead of publishing a universal percentage, we ran controlled text-image fixtures through the same Tesseract command used by NetroDoc's Image to Text tool and recorded the character-level result for each supported language.

Controlled evidence · 2026-09-09

What the measured run showed

These figures describe the controlled fixture documented for this guide. They are evidence for the tested case, not a universal performance guarantee.

OCR engineTesseract 5.3.4
Page modePSM 3
Languages tested7
Latin clean fixtures100% in this test
Arabic fixture81.82% in this test
OCR page limit25 pages

The test method

We rendered controlled text images and processed them with Tesseract using the same language selection and --psm 3 page-segmentation mode used by NetroDoc's Image to Text path. The test also created a degraded version with downscaling, blur, and slight rotation.

Character accuracy was calculated against the expected text after normalizing case and whitespace. These values describe the fixtures we created on 9 September 2026; they are not promised accuracy rates for arbitrary scans, handwriting, receipts, or camera photos.

  • English: 100% clean and 100% degraded in this fixture.
  • Dutch: 100% clean and 100% degraded in this fixture.
  • German: 100% clean and 100% degraded in this fixture.
  • French: 100% clean and 100% degraded in this fixture.
  • Spanish: 100% clean and 100% degraded in this fixture.
  • Portuguese: 100% clean and 100% degraded in this fixture.
  • Arabic: 81.82% clean and 81.82% degraded in this fixture.

Why the 100% results must not be generalized

The Latin-script fixtures contained a single clear line in a known font on a clean background. Even the degraded variants were controlled synthetic images, not badly lit phone photos or old photocopies. A perfect score here proves that the tested path can recover those samples; it does not prove that every document in the same language will score 100%.

Real OCR difficulty rises with low resolution, compression artifacts, skew, shadows, mixed columns, decorative fonts, handwriting, curved pages, faint ink, and text placed over images. A language pack can be installed and working while a particular page is still hard to segment correctly.

What the Arabic result tells us

The Arabic fixture did return readable Arabic text, but character-level accuracy was lower than the Latin fixtures. The recorded output included errors around digits and one word. That is exactly the kind of result a useful OCR guide should expose instead of hiding behind a generic accuracy claim.

For Arabic documents, review names, dates, numbers, and short fields carefully. Those details can matter more than whether a paragraph is broadly readable.

  • Choose Arabic explicitly when the page is primarily Arabic.
  • Use a straight, high-resolution scan with strong foreground/background contrast.
  • Verify dates, amounts, identification numbers, and names against the image.
  • Do not treat OCR output as authoritative transcription without review.

How to improve a difficult scan before retrying

OCR works best when the engine can first identify where text lines are. Cropping away large irrelevant borders, correcting rotation, increasing effective resolution, and avoiding heavy JPEG damage can help page segmentation before character recognition even begins.

If the source is a PDF scan, searchable-PDF OCR is useful when you want to preserve the page image while adding a text layer. Image to Text is better when the extracted plain text itself is the deliverable.

Current NetroDoc OCR scope

The tested OCR language set is English, Dutch, Arabic, German, French, Spanish, and Portuguese. Italian is a website language but is not currently an OCR language option, so the site should not imply otherwise.

The configured OCR limit in the tested environment is 25 pages. OCR jobs are temporary rather than permanent document storage.

Temporary processing and privacy

Our lifecycle test verified that the OCR result folder was removed after the download cleanup ran and that expired OCR folders were removed. The configured OCR job TTL was 3,600 seconds in the tested environment.

Read the Privacy Policy →

Troubleshooting

OCR returns little or no text

Check that the correct language is selected and that the page contains a readable, well-oriented text region. Large borders or unusual page structure can also hurt segmentation.

Most words are right but numbers are wrong

Manually verify dates, amounts, IDs, and other short high-value fields. Our Arabic fixture specifically showed digit errors.

A scanned PDF is not editable

Use OCR first to add or extract a text layer. PDF-to-Word alone does not OCR image-only pages.

Method note

NetroDoc publishes measured claims only after a controlled run completes successfully. Untested cases are kept out of the conclusion. The evidence run for this publication wave used the same production runtime components while operating on an isolated candidate and temporary server paths.

Back to all Guides →