How to Translate a Scanned PDF (the Honest Way)

If your PDF is a scan, no translation tool can translate it directly — including ours. A scanned page is a picture of text, not text. There is nothing to send to a translation engine until you run optical character recognition (OCR) to create a text layer.

So the workflow is two steps, not one:

1. OCR the scan with a dedicated tool to produce a searchable PDF.
2. Translate that searchable PDF with a layout-preserving translator.

The rest of this article is how to do each step without wasting a day on it.

First: is your PDF actually scanned?

Open it and try to select a sentence with your cursor.

• Text highlights word by word — you have a real text layer. Skip straight to translating.
• Nothing selects, or a blue box covers the whole page — that page is an image. You need OCR.

Two more quick tells:

• Search fails. Press Ctrl+F (Cmd+F on Mac) and search for a word you can plainly see. No hits means no text layer.
• The file is oddly large. A 12-page text document is often well under 1 MB. A 12-page scan is frequently 10 to 40 MB, because every page is a photograph.

Watch out for mixed documents. A contract exported from a word processor with a scanned signature page appended is text on pages 1 to 9 and an image on page 10. Check a few pages, not just the first one.

Step 1: run OCR to create a searchable PDF

You want output that is still a PDF, with the original page images intact and an invisible text layer added underneath. That format is usually called a searchable PDF. Avoid tools that dump the recognized text into a plain document — that throws away the layout you are trying to keep.

Reasonable options, described in general terms:

• Adobe Acrobat includes a built-in OCR feature for scanned documents. It is a paid, commercial product, and it is the path of least resistance if your organization already licenses it.
• OCRmyPDF is a free, open-source command-line tool that adds a text layer to an existing PDF while leaving the page images alone. It is built on Tesseract, the long-running open-source OCR engine. This combination is the standard free answer, and it handles batches well.
• Cloud document services can often recognize text in an uploaded PDF. These are useful for a quick quality check, but many of them convert the file into an editable document and discard the original page layout in the process.
• Built-in OS features increasingly recognize text inside images — enough to spot-check whether a scan is legible at all, though not usually enough to produce a whole searchable PDF.

Whichever you use, two settings matter more than the tool you pick:

• Set the OCR language to the source language of the document. Recognizing German text with an English model produces confident nonsense.
• Scan or rescan at around 300 DPI if you control the source. Below roughly 200 DPI, accuracy falls off sharply and no amount of post-processing recovers it.

Step 2: check the OCR before you translate

This is the step everyone skips, and it is the one that decides whether the final translation is usable. OCR errors do not stay put — they compound. A misread word becomes a mistranslated word, and unlike a typo in the original, nobody downstream can tell it was ever a scanning artifact.

Open the OCR'd PDF and run this checklist:

• Select and copy a full paragraph, then paste it somewhere you can read it plainly. Are the words right? Are the line breaks sane?
• Search for three domain-specific terms — a product name, a chemical, a legal term. These are what OCR gets wrong most often, and they are what your readers will notice.
• Check the numbers. Confusions between 0 and O, 1 and l, 5 and S, and 8 and B are the classic failures, and in an invoice or a lab report they are the whole document.
• Check hyphenation at line ends. Words split across lines often come back as two fragments, which translate as two fragments.
• Look at every table. Tables are the hardest thing in OCR. Cell boundaries get lost, columns merge, and a merged column produces a translated table that reads correctly line by line and is entirely wrong.

If the copied text is garbage, stop. Re-scan at higher DPI, deskew the pages, or correct the OCR language and try again. Translating bad OCR just launders the errors into a second language.

Step 3: translate the searchable PDF without flattening it

Now you have a real text layer — but you are not out of the woods, because the next common failure has nothing to do with scanning.

Many translation pipelines take a PDF and flatten it to a plain text stream before translating. On anything with columns, that destroys reading order. Here is real output from running a standard text extractor over page one of a two-column government fact sheet:

S
ince the early 1990's, the
magnitude and location of an
Not long ago, the fi rst thing that most
people did after feeling an earthquake was to

Look at what happened. The drop cap "S" is severed from "ince". Two different columns are interleaved mid-sentence. The "fi" ligature broke into "fi rst". The reading order was wrong before a single word was translated — so the translation is a faithful rendering of scrambled text.

The example above comes from a born-digital PDF, so an OCR text layer will not reproduce that exact drop-cap and ligature signature. But column interleaving is the same failure, and it is the one that will bite you: reports and journals are frequently multi-column, and that is precisely the material people scan.

PDFPal takes the other approach: it separates the visual layout from the text, translates the text with formulas masked, then re-anchors the translation back onto the original page. Columns, tables, figure placement, and math stay where they were, and you get both a translated-only PDF and a bilingual side-by-side version. You can see real before-and-after pairs from six public-domain documents at examples. Sign-up includes 20 free pages with no credit card, which is enough to test a chapter of your OCR'd scan before committing to the whole thing.

What PDFPal does and does not do with scans

Being specific, because this is where trust is won or lost:

• There is no OCR in PDFPal. We do not perform text recognition, and we are not going to imply otherwise.
• If you upload a scan anyway, the job detects that the pages have no text layer, fails with a clear message, and refunds your credits automatically. You are not charged for an impossible job.
• There is a "scanned-PDF mode" toggle in the translate screen. It improves rendering on image-heavy pages — forcing clean black text on white. It does not add text recognition. It only helps once a text layer already exists.
• Input must be a PDF, up to 100 MB. If your scan is larger, split it.

What OCR will not fix

Set expectations before you invest the time:

• Handwriting. Standard document OCR is built for printed type. Handwritten notes, margin annotations, and signatures will not come through usably. Plan to transcribe those by hand.
• Very low resolution. Faxes, photocopies of photocopies, and phone snapshots taken at an angle produce error rates that make review slower than retyping.
• Stamps, watermarks, and marginalia overlapping body text confuse recognition and inject junk words into the middle of otherwise clean sentences.
• Dense tables and forms, as above — budget manual review time for these regardless of how good the rest looks.

The short checklist

1. Try to select text. Nothing selects means it is a scan.
2. OCR it into a searchable PDF with the correct source language, from a 300 DPI source if possible.
3. Copy a paragraph and verify it. Check terms, numbers, and tables.
4. Fix and re-OCR if the text is wrong. Do not translate bad OCR.
5. Translate the OCR'd PDF with a tool that preserves layout, and review the tables in the output.

It is more work than dragging a file into a box. But the alternative is a translated document that looks finished and is quietly wrong — and that costs far more to discover later. When your PDF does have a text layer, translating it is the easy part; see pricing for how credits work.