The short version

There is no text in a scanned PDF to extract. OCR reads the picture and produces its best guess at the words, which is very good on a clean 300 DPI scan and unreliable on a phone photo.

Expect to lose layout. Columns, tables and headers are structure, and text-only output keeps words rather than structure.

The word extract sets the wrong expectation. Extracting text from a born-digital PDF genuinely is extraction: the characters are in the file and a tool reads them out. A scan contains no characters at all, only an image that happens to depict them.

Extraction versus recognition

This distinction explains why results vary so wildly between two documents that look equally readable on screen.

AnatomyWhy this is recognition, not extraction
Imagea picture of a pageLayoutfind text regionsSegmentsplit into charactersR e a dClassifymatch each shapea mistake at any stage survives into the text
There is no text in the file to extract. Every word has to be read out of a picture.
Born-digital PDF

The text is really there

  • Characters stored as codes with a font reference
  • Extraction is exact, with no guessing involved
  • Works instantly regardless of page appearance
  • Reading order usually preserved
  • No proofreading needed
Scanned PDF

The text has to be guessed

  • Pages are photographs of printed characters
  • OCR classifies each shape and may be wrong
  • Accuracy depends on the scan, not the software
  • Layout analysis can misread column order
  • Output always needs checking
Check which one you have first
Open the PDF and try to select a line of text with your cursor. If individual words highlight, the text is real and you can extract it directly with no OCR involved. If the whole page highlights as one block, or nothing does, it is a scan and recognition is the only route.

What you lose

OCR produces words. Everything that made those words a document is a separate layer of information that plain text output cannot carry.

Text-only outputKept or lost
  • The words themselvesKept
  • Paragraph breaksUsually kept
  • Bold and italicLost
  • Table structureLost
  • Multi-column orderOften scrambled
  • Images and figuresLost entirely

Tables are the painful case. The engine may read every cell correctly and still hand you a stream of words with no indication of which column each belonged to. A financial table becomes a list of numbers in an order that is difficult to reconstruct, and for anything with more than a few rows it is often faster to retype it than to repair the output.

Multi-column pages can come out interleaved
Layout analysis decides where the columns are before any reading happens. When it gets that wrong on a newspaper-style page, it reads straight across both columns, producing sentences that alternate between two unrelated topics. The individual words are correct and the text is useless. Check the beginning of the output before trusting the rest.

The mistakes OCR makes

Recognition errors are not random. They cluster around characters whose printed shapes are genuinely similar, which means you can search for them directly.

Confusion pairs worth searchingCommon substitutions
  • rn read as mand the reverse
  • l, 1 and Iinterchangeable in many fonts
  • 0 and Onear identical in some faces
  • cl read as dwhen letters touch
  • 5 and Sespecially at low resolution
  • 8 and Bin faded or thick print
Numbers are riskier than words
The language model that fixes ambiguous letters works by checking whether the result is a real word. Numbers have no dictionary, so nothing corrects a misread digit. An invoice total, an account number or a date can come back confidently wrong with no signal that anything happened. Always check figures against the image rather than trusting the text.

Proofreading efficiently

Reading the whole output against the original is slow and mostly wasted effort, since the great majority of it will be correct. Targeted checks find more errors in less time.

1

Check every number against the page

Totals, dates, reference numbers, quantities. This is where errors are both most likely to survive and most damaging, because nothing corrects them automatically.

2

Search for the confusion pairs

Look for stray 1 inside words and l inside numbers. A handful of searches covers the most common substitutions across the whole document.

3

Read proper nouns carefully

Names, places and companies are not in the dictionary, so the language model cannot help and may actively harm by correcting them towards real words.

Getting a better result

Almost all accuracy is decided before the software runs, which means the highest-value fix is usually to go back to the paper.

Rescanning at 300 DPI, in greyscale, straight on the glass, with the scanner's deskew enabled, will beat any amount of reprocessing of a poor image. If the original is not available, cleaning up contrast and straightening the existing image before recognition still helps, though it cannot recreate detail that was never captured.

Choose the right language before running
The language model is what resolves genuinely ambiguous shapes, and it can only do that if it knows which language it is reading. Pointing it at the wrong one turns a useful correction stage into a source of confident errors. For a document mixing languages, run it in whichever dominates and expect to fix the rest.

Run recognition in your browser

The scan is read on your own machine, so neither the file nor the words in it are sent anywhere.

Open OCR PDF

Common questions

Working with what comes back.

01Why can I not just copy the text out of my scan?

Because there is no text in it. The pages are photographs, so there is nothing for a copy operation to find. OCR has to read the image and produce text before anything can be copied.

02How accurate is OCR?

Very high on a clean, straight 300 DPI scan of ordinary printed text. Considerably lower on low resolution scans, angled phone photos, faded originals and unusual typefaces. The scan quality matters far more than which software you use.

03Why did my table come out as a jumble?

Because OCR recognises characters, not structure. It can read every cell correctly and still have no way to record which column each belonged to. For anything but the simplest tables, retyping is often faster than repairing the output.

04Why is my two-column page interleaved?

Layout analysis misjudged the columns and read straight across the page instead of down each one. The words are correct and the order is not. Check the first few lines of output before relying on the rest.

05Which errors should I look for first?

Numbers, because no dictionary corrects them, and proper nouns, because the language model cannot help and sometimes makes them worse. Then search for the classic confusion pairs such as rn against m, and l against 1.

06Can I improve the result without rescanning?

Somewhat. Straightening the page and improving contrast before recognition helps. But detail that was never captured cannot be recovered, so if the original document is available, rescanning at 300 DPI in greyscale is by far the most effective fix.