OCR (Optical Character Recognition)
Turning pictures of text into text — solved for clean documents, still genuinely hard for everything else.
Reviewed July 11, 2026Stable
When not to use it
- When the text is already digital. Extracting from a PDF that has a text layer beats re-reading the pixels — check before you build.
- When layout is the point. Reading every character off an invoice without knowing which number is the total is not a result. That's document understanding.
- On low-resolution images, hoping. Below roughly 300 DPI, accuracy falls away and upscaling doesn't restore information that was never captured.
Reach for something else instead
- PDF text extraction when there's a text layer — exact, instant, free.
- Multimodal models when layout and meaning matter more than character-perfect transcription, and you can tolerate the cost and the hallucination risk.
- Structured data at source. Often the real answer is asking for a CSV rather than reading a picture of one.
Further reading
- Graves et al. (2006), Connectionist Temporal Classification — the alignment trick underneath most text recognition.
- Shi, Bai & Yao (2015), An End-to-End Trainable Neural Network for Image-based Sequence Recognition — CRNN, the architecture most engines still resemble.
- Xu et al. (2020), LayoutLM: Pre-training of Text and Layout for Document Image Understanding — reading the page, not just the characters.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Testing on clean scans and deploying to phone photos. Perspective, shadow, and focus are a different problem entirely.
- Trusting a multimodal model's extracted numbers without validation. Unlike OCR, it fails silently with a plausible answer.
- Reporting character error rate as if it captures business impact. One wrong digit in a total barely moves CER and is a complete failure.
At a glance
FieldComputer Vision
Core ideapictures of text into text
Solved forclean printed scans
Hard forhandwriting, photos, layout
Floor~300 DPI
DifficultyIntermediate
Often compared with
OCR vs. multimodal models — transcribing characters reliably vs. understanding a page and occasionally inventing it.
Where this sits
A destination. 10 concepts lead here, and nothing in the corpus depends on it.
7Levelsteps in
10Needs firstconcepts
0Opens upnothing further
1Areastays here
Learn these firstImage Classification
Computed from the prerequisite graph, not assigned. How this works