Common OCR mistakes and how to fix them

OCR is probabilistic. Even strong models misread characters, merge lines, or invent punctuation when the pixels are ambiguous. Knowing the common failure patterns makes cleanup fast — and tells you when to recapture instead of editing.

Character confusions

Seen	Often meant	Notes
O / 0	letter O vs zero	Critical in IDs and amounts
l / 1 / I	el / one / eye	Sans-serif fonts blur these
S / 5		Weak in noisy scans
B / 8		Soft focus
rn / m		“modern” → “rnodern” classics

Fix: Search the output for impossible tokens in context (a price of O9.99, an order id with letters where only digits are allowed).

Broken reading order

Multi-column layouts may read across columns: line 1 left + line 1 right as one sentence. Magazines and academic papers are frequent offenders.

Fix: Crop each column to its own image, convert separately, then concatenate. Or convert the full page and split manually if the mistake is rare.

Tables that “melt”

Missing gridlines cause cells to merge. Extra stains create phantom columns.

Fix: Recapture with higher contrast; try a flatbed scan; convert table-only crops. Verify with the tables guide.

Hyphenation and line breaks

Scanned books hyphenate at line ends (infor- / mation). OCR may keep the hyphen or join wrong.

Fix: Search for - followed by a newline pattern in the Markdown and rejoin words.

Headers, footers, and stamps

Notaries, “CONFIDENTIAL” watermarks, and page numbers pollute body text.

Fix: Delete repeating lines in bulk after conversion. For watermarks across letters, recapture is rarely possible — manual correction of hit words may be required.

Low-contrast and colored text

Yellow text on white, or gray on gray, drops characters entirely.

Fix: Adjust capture lighting; increase display zoom for screenshots; use light mode.

Compression artifacts

Heavy JPEG compression around text creates ringing that looks like extra punctuation.

Fix: Use original screenshots (PNG) or higher-quality transfers. Avoid screenshot-of-screenshot chains.

When to re-run vs edit

Re-run (new capture or crop) when:

• Whole regions are missing
• Skew is severe
• Glare covers a paragraph

Edit the Markdown when:

• Only a handful of characters are wrong
• Structure is good but a header repeated
• One table cell failed

A cleanup pass that scales

For a long document:

1. Fix global issues (hyphenation, repeating headers).
2. Skim headings for outline sanity.
3. Spot-check numbers and names.
4. Only then format for publishing.

Do not start bolding titles while digits are still wrong.

Preventing mistakes upstream

Most “OCR bugs” are capture bugs. Revisit:

• Photograph documents for better OCR
• Extract text from screenshots
• Convert image to text

Still stuck?

If a page repeatedly fails, the content may be outside comfortable OCR territory (heavy handwriting, artistic lettering, extreme damage). In those cases, partial OCR plus human transcription of the hard lines is the honest workflow.

You can also contact us if you think the tool itself is misbehaving on a clear, high-quality file — include a description of what you tried (please do not send highly sensitive documents over email).