Tips & Tutorials

A Quick Guide to What OCR Can and Can't Do

A Quick Guide to What OCR Can and Can't Do

Optical character recognition, OCR, is the technology behind turning a scanned document or a photo of text into actual selectable, searchable, copyable text. It is genuinely useful and, for the right kind of source material, impressively accurate — but it also has real, predictable limits that are worth understanding before you rely on it for something important, rather than discovering them by trial and error on a document that actually matters.

What OCR is actually doing

At a basic level, OCR analyses an image, identifies regions that look like text, and matches the shapes it finds against known character patterns to decide what each one most likely is. It is fundamentally a pattern-recognition problem, not genuine reading comprehension — the software has no understanding of what the words mean, only what shapes most closely resemble which letters, which is exactly why context-dependent mistakes (correctly shaped but wrong-in-context words) are a known category of OCR error.

What OCR handles well

  • Clean, printed text — a straight, well-lit, high-resolution scan or photo of printed text in a standard font is where OCR performs at its best, often with accuracy close to perfect.
  • Standard fonts at a reasonable size — common typefaces at normal body-text size, not tiny print or unusual decorative fonts, recognise most reliably.
  • High contrast between text and background — black text on a plain white background is the easiest case; anything that reduces that contrast makes the job harder.
  • Straight, unrotated text — text that runs horizontally, the way it was printed, rather than at an angle.

Where OCR genuinely struggles

  • Handwriting — handwritten text varies enormously between individuals in ways printed fonts simply do not, and general-purpose OCR built for printed text is much less reliable on it. Results vary a lot, and should be checked carefully rather than trusted outright.
  • Low light or blurry photos — a photo taken in dim light, or one that is slightly out of focus, loses exactly the crisp character edges OCR depends on to tell similar-looking letters apart.
  • Text at an angle — a crooked scan or an angled photo distorts character shapes in a way that significantly increases misreads, even when the text itself is otherwise clear.
  • Unusual or decorative fonts — stylised typefaces, especially anything resembling handwriting or heavy ornamentation, confuse recognition models trained primarily on standard printed fonts.
  • Very small text — fine print, footnotes, and anything below a certain size relative to the scan's resolution loses the detail OCR needs.

What OCR does not do, even when it works well

OCR recognises and extracts text; it does not understand or summarise it, does not verify that numbers add up correctly, and does not know when something looks unusual enough to flag for a human to double-check. It also does not preserve complex formatting reliably — tables, multi-column layouts, and text wrapped around images often come out in a different reading order than the original layout intended, since the recognition process is working through the page roughly left-to-right, top-to-bottom, without necessarily understanding the visual structure a human reader would intuitively follow.

How to get noticeably better results from the same document

  1. Scan rather than photograph where possible — a flatbed scanner produces flatter, more evenly lit, straighter images than a handheld phone photo.
  2. If photographing, use even lighting, hold the camera directly above the page rather than at an angle, and avoid any shadow falling across the text.
  3. Scan at a reasonable resolution — too low and fine detail is lost; going far beyond around 300 DPI rarely helps accuracy further and just produces a larger file.
  4. Always review the output against the original for anything where accuracy genuinely matters — numbers, names, dates — rather than trusting it blindly.

Trying it yourself

Our own OCR PDF tool follows this same general pattern: it reads printed English text well under good scan conditions, and like any OCR system, gets less reliable on handwriting, low light, or an angled source. Uploading a test page you already know the content of is a good way to get a feel for how it handles your specific kind of documents before relying on it for something you have not already read closely yourself.

A realistic way to think about OCR accuracy

Treat OCR as a very fast first draft of extracting text, not a guaranteed-accurate final transcript. For casual use — searching through a pile of old scanned notes, say — near-perfect accuracy is rarely necessary, since you are looking for approximate matches, not relying on every character. For anything where a single wrong character actually matters — a legal document, a medical record, financial figures — always verify the specific fields that carry real consequences, even if the rest of the page is left unchecked.

Frequently asked questions

Does OCR work on any language?

OCR systems support specific languages they are trained on. Our own tool currently recognises English text.

Can OCR read text inside an image embedded in a PDF, not just the page itself?

Yes, if the embedded image contains visible text, OCR processes it the same way as any other page content, since it works from the rendered page image.

Why did OCR get some words right and others completely wrong on the same page?

Local factors matter — a slightly blurrier patch, a faint fold crease, or a smaller font size in one part of the page can affect just that region's accuracy while the rest of the page reads fine.

Is there a way to improve accuracy after the fact, without re-scanning?

Not meaningfully — OCR accuracy is determined by the quality of the image it is given. If results are poor, a cleaner re-scan or re-photograph is the most effective fix.

Should I trust OCR output for a document I will submit somewhere official?

Always review it first, particularly any numbers, names or dates, since OCR mistakes look like plausible text rather than obvious errors.

A simple mental model to keep in mind

It helps to think of OCR as being only as good as the clearest version of the text a human could read from the same image — if a person would need to squint, guess, or give up on a particular word, OCR is working from the same limited information and will likely do the same. This mental model makes it much easier to predict, before you even run a document through OCR, roughly how reliable the result is likely to be.

Related guides

← Back to all posts

Keep reading