How OCR works, and why it still makes mistakes
OCR is the step that turns a photographed or scanned page into words you can search, copy and edit. It is remarkably good today, and still a set of educated guesses that deserve a proofread.
- Published
- CategoryAI & Technology
- Reading time4 min read

What OCR actually does
A scanned page or a phone photo of a document is just a grid of coloured pixels. Your eyes see letters; the computer sees light and dark dots. Optical character recognition (OCR) is software that looks at those pixels and produces real characters, so that "Total: 1,250.00" becomes text you can search, copy into a spreadsheet or edit in Word.
Without OCR, converting a scan to another document format only moves the picture around. That is why a scanned PDF converted to Word usually gives you page images rather than editable text, as explained in our guide to why scanned PDFs are not editable.
The main steps inside an OCR engine
Different engines work differently, but most follow a similar sequence:
- Clean-up. The image is straightened (deskewed), converted to high-contrast black and white, and cleaned of specks and shadows.
- Layout analysis. The engine finds blocks of text, columns, tables and images, and decides the reading order.
- Line and character recognition. Older engines matched individual character shapes against patterns. Modern engines typically use neural networks that read whole lines or words at once, which copes better with joined letters and unusual fonts.
- Language modelling. The engine uses knowledge of the language, such as common words and spelling, to choose between similar-looking options. "rn" versus "m" is often settled this way.
- Output. The result can be plain text, a searchable PDF (the original image with an invisible text layer on top), or a document format.
The searchable PDF is worth knowing about: the page still looks exactly like the scan, but you can now select and search words, and a PDF-to-Word converter has text to work with.
Why OCR makes mistakes
Recognition is interpretation, so errors are normal. The most frequent causes:
- Low resolution. Small text photographed from far away has too few pixels per letter. For ordinary printed documents, around 300 dpi scanning, or a sharp phone photo filling the frame, gives the engine much more to work with.
- Blur, shadows and curved pages. A book photographed at an angle has curved lines and uneven lighting.
- Unusual fonts, handwriting and stamps. Printed text is far easier than handwriting. Signatures, stamps and decorative fonts often produce nonsense.
- Complex layouts. Multi-column pages, sidebars and tables can be read in the wrong order even when every word is recognised.
- Mixed languages and scripts. A page combining Arabic and English, or a technical page full of codes, needs an engine configured for both.
The errors that matter most
Most OCR errors are easy to spot because they produce strange words. The dangerous ones look plausible: 5 and 6, 1 and 7, 0 and O, a missing decimal point, a minus sign read as a dash. In invoices, bank statements and medical or legal documents, a single wrong digit can change the meaning completely.
How to get better OCR results
- Start with the best image you can. Flat page, even daylight, phone held parallel to the paper, no zoom. Retaking a photo is faster than correcting a bad result.
- Rotate and crop first. Sideways or upside-down pages confuse layout analysis. If your scan is a PDF, fix orientation before running OCR.
- Do not shrink the image too much. Resizing a photo to save space before OCR removes the detail the engine needs. If you must resize for an upload limit, do it after recognition. The Image Resizer shows the exact pixel dimensions you end up with.
- Choose the right language. Most OCR applications let you select the document language. This improves accuracy noticeably.
- Process difficult pages separately. Tables, forms and handwritten notes often need manual retyping anyway.
How to check OCR output
- Search the result for a few words you can see on the page. If they are found, the text layer exists.
- Compare every number in totals, dates, account references and quantities with the image.
- Read headings and the first line of each paragraph to catch reading-order problems.
- For important documents, keep the original image or scanned PDF next to the recognised version.
Where OCR fits with Convertios tools
Convertios does not currently offer an OCR tool, and its PDF to Word converter does not perform OCR. The practical workflow is to run OCR in an application that offers it, such as a PDF editor, an office suite or a phone scanning app, and then use Convertios for the next step: converting the searchable PDF to Word, or exporting individual pages as JPG when you need images instead of text.
OCR turns pictures of text into text, but it guesses. Treat every OCR result as a draft until the numbers and names have been checked against the original.
Published by Convertios.


