OCR (Optical Character Recognition) — AI Glossary

Reference details

Related terms

• Intelligent Document Processing
• Document Q&A
• DocVQA

How it works

1. The image is pre-processed — deskewed, denoised, binarised — because recognition accuracy depends heavily on input quality.
2. Text regions are detected before being read, so the engine knows where characters are rather than scanning uniformly.
3. Characters or lines are recognised by a model, and a language model over the output corrects implausible sequences.
4. A confidence is emitted per token or region, which is the signal that lets low-confidence output be routed rather than trusted.
5. Vision-language models can do this end to end without the staged pipeline, trading the per-region confidence signal for robustness on messy layouts.

Example: Why character accuracy is the wrong headline

At 99% character accuracy, a 10-character account number is correct 0.99¹⁰ ≈ 90% of the time — so one in ten documents carries a wrong identifier. The page as a whole, at 2,000 characters, is essentially certain to contain errors somewhere. Character accuracy sounds excellent and the field-level number is what matters: 90% on the field that keys a payment is not a passing grade, which is why per-field confidence and validation exist.

Commonly confused with

Used as a synonym for document processing, which contains it. This converts images of text into characters; the surrounding pipeline handles layout, classification, extraction and validation, and most failures live there. A native digital PDF needs no recognition at all and can still produce scrambled text, because the failure was reading order rather than characters. Checking which of the two is at fault takes one look at the extracted text and saves tuning the wrong stage.

When to use it

Reach for it when:

• Source documents are scans or photographs with no extractable text layer.
• Per-region confidence is needed to route uncertain output to review.
• Explaining why a scanned corpus needs a step a digital one does not.

Reach for something else when:

• Documents have a text layer, where extraction is direct and recognition adds error.
• As the explanation for a layout failure on a digital PDF, where it was never involved.
• Reporting character accuracy as the quality figure when field accuracy is what matters.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to OCR (Optical Character Recognition) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path