Why Is My OCR Text Full of Errors? Scan Quality, Cleanup, and Verification

You ran OCR on a scan, and the text layer came out wrong: names misspelled, digits transposed, whole paragraphs reading as plausible-looking gibberish. The instinct is to call the engine inaccurate. In practice the engine is reading exactly what is in the image, and the image is what needs fixing first. This guide covers the scan defects that produce most OCR errors, the order to clean them up in, and the check that tells you the result is safe to rely on.

Why is my OCR text inaccurate?

OCR errors almost always trace back to the image, not the engine: low resolution, skewed pages, uneven shadows or yellowed paper, content touching the text, and the wrong language assumption are the five causes behind nearly every bad result. Fix the image first, then run OCR, then verify — in that order. Re-running OCR on the same broken scan returns the same broken text every time.

What you see in the text layer	The defect in the image	The fix to apply first
rn reads as m, digits confused with each other	Low resolution	Re-capture the page at a proper scan resolution
Lines from two paragraphs interleaved	Skewed or rotated page	Straighten with the rotation tool, then the deskew pass
Whole lines or bands missing	Shadow band across the page	Run the scan enhancement pass
Characters lost where the paper is yellowed or grey	Low contrast background	Scan enhancement, which flattens the background
Text corrupted only under a stamp or a border	Content overlapping the glyphs	Crop the border, and accept stamp areas as unrecoverable
Confident nonsense in a non-Latin script	Wrong language assumption	Confirm the document’s language and scan the settings you ran with

What are the possible causes?

• Low resolution. A page captured at well below a proper scan resolution does not contain enough pixels per character for any engine to distinguish similar glyphs — rn becomes m, 3 becomes 8, cl becomes d. Text photographed at a distance from a phone is the usual source. As a working example, the resolution most scanning guidance recommends for readable text documents is 300 DPI, and recovery from half of that is limited no matter which engine you use. Real example: a photographed two-page contract produces a text layer in which every rn pair reads as m — the engine is not guessing, it is resolving genuinely ambiguous pixels.
• Skewed or rotated pages. A page tilted even a few degrees breaks line detection: reading order splits, columns merge, and the text layer mixes lines from two paragraphs. Pages fed through a scanner at an angle, or photos taken holding a phone by hand, are the common sources.
• Shadows, uneven lighting, and yellowed paper. A shadow band across the page turns into a black stripe after thresholding, taking the lines under it with it. Yellowed or grey background paper reduces contrast so the binarization step cannot cleanly separate glyphs from background.
• Content touching or overlapping the text. Red stamps and seals over the text, underlines, table borders merged into glyph strokes, and punched holes at the edge all corrupt the character shapes the engine reads.
• The wrong language or character set. OCR runs with a language model; a document read with the wrong one produces confident nonsense, particularly with mixed Latin and CJK content.

How do I fix it, step by step?

1. Look at the source page, not the text layer. Convert the suspect pages to images with PDF to images and look at them at full size. You can see low resolution, skew, shadows, and stamp overlap directly — this one look tells you which of the causes above you are fighting. If the document is long, check the first page, the last page, and one page from the middle: defects from a physical capture are not evenly distributed, and a clean first page does not guarantee a clean middle.
2. Straighten the pages. If pages are rotated, fix them with rotate specific PDF pages before anything else. If the tilt is only a few degrees, the enhancement tool in the next step handles deskew as part of its pass.
3. Run the scan enhancement pass. Enhance a scan applies local adaptive binarization, removes uneven shadows, flattens yellowed backgrounds, sharpens glyphs, and preserves coloured stamps — all in your browser. This single pass addresses the shadow, background, and contrast causes at once and is usually where the largest accuracy gain comes from. If black edges from the scanner glass or punched holes remain, crop them with crop PDF pages. Do this before OCR, not after: enhancing a file that already carries a text layer does not improve the layer, because the layer was built from the unenhanced image.
4. Run OCR. Use OCR a PDF: it runs a WASM engine locally, no upload, and produces a dual-layer output with an invisible text layer laid over the original page image, so search and copy work while the visual page stays untouched.
5. Verify the text layer before you rely on it. Check with check PDF text layer — it reports page by page which pages have a text layer, so a silent gap on one page shows up immediately. Then open the file in view a PDF, search for three names, three numbers, and one uncommon word from the document, and confirm each lands on the right page. Deep verification guidance for the whole archive is in Make a PDF Searchable.
6. Re-run only the pages that are still wrong. If the text layer is accurate on twenty pages and wrong on three, fix the cause on those three pages — usually resolution or a shadow — and re-run OCR on the affected range rather than the whole file. Re-running everything wastes nothing but your time; re-running a still-broken page wastes the result.

How do I prevent it next time?

• Capture at a real scan resolution. For text documents, scan rather than photograph where possible — a flat, even scan at a proper resolution is the single biggest predictor of OCR accuracy.
• Scan flat, on clean glass, with the page straight. Most skew and shadow defects come from the physical capture step and never reach the file if the capture is clean.
• Enhance before you OCR, once. Running the enhancement pass before OCR costs one extra step and removes the shadow and background causes that engines cannot handle on their own.
• Verify the text layer, don’t assume it. A page-by-page text-layer check takes seconds and catches the silent failure mode — an OCR run that simply skipped some pages.
• Keep handwriting out of OCR-bound workflows. Handwritten annotations and signatures are not reliably machine-readable glyph sequences; treat them as images, and plan the workflow around that instead of hoping.

Frequently Asked Questions

Where does the OCR run — is my scan uploaded? No. OCR a PDF runs a WASM engine inside your browser tab. The scan never leaves your device, and no account is required.

Why is accuracy worse on some pages than others? Because the defects are not evenly distributed across a capture. A shadow from a curled corner, a page fed at a different angle, or a fingerprint on the glass degrades exactly the pages it touched. The page-by-page text-layer check is what makes the unevenness visible.

Can OCR read handwriting? Not reliably. OCR resolves printed glyphs; handwriting varies too much per person to resolve in general. If a document is mostly handwritten, transcription by a human is the honest path, and the PDF tools here stay useful for the structure around it.

Why is the text layer invisible in the output? The dual-layer design lays an invisible text layer over the original page image, so search, selection, and copy-paste work while the page still looks exactly like the scan. If you want a normal-looking text document rather than a scan, follow the conversion path in Why Is My PDF Conversion Output Wrong?.

The OCR text is still wrong after cleanup. Now what? The remaining cause is resolution: if the source capture lacks the pixels, no engine recovers the characters. Re-capture the page properly — that fix is upstream of any tool.

PDFLean tools for this job

All in-browser, all free, no account: OCR a PDF, enhance a scan, check PDF text layer, PDF to images, rotate specific pages, crop PDF pages, view a PDF, scan to PDF.

Related reading

• Make a PDF Searchable — the full workflow when the goal is a searchable archive, not one file
• Why Is My PDF Conversion Output Wrong? — what to do once the text layer exists and the conversion still misbehaves
• OCR Text Layer — the one-paragraph model of what a dual-layer PDF actually carries