How to OCR a scanned PDF privately

A scanned PDF often contains pictures of pages rather than real text. You can see the words, but you cannot search, select or copy them because the computer only sees pixels. Optical character recognition, usually shortened to OCR, analyzes those pixels and predicts the characters and word order. The result is useful, but it is not a perfect transcription.

Privacy matters because scans frequently contain identity documents, invoices, medical correspondence and signed forms. An on-device OCR engine downloads its language model and performs recognition inside the browser. The page images do not need to be sent to a remote recognition service. The first model download may be large, so load it on a trusted connection before working offline.

OCR quality begins with the source image. Pages should be upright, evenly lit and large enough that small characters have clear edges. Shadows along a book spine, patterned backgrounds and compression blocks can be interpreted as punctuation. If you control the scan, place the page flat, avoid glare and include the full margins. A slightly larger clean image is usually better than an aggressively compressed one.

Rotate sideways pages before OCR. Crop away a large desk or background area so the model focuses on the document, but do not trim page numbers or footnotes. For a faint page, improving contrast can help; for a photograph or shaded form, extreme black-and-white conversion may erase meaningful marks. Keep the original scan so every cleanup choice is reversible.

Choose the language that dominates the document. Language models use spelling and character patterns to separate similar shapes, so the wrong choice can turn accented letters into punctuation or confuse common word endings. A bilingual document may require separate passes or closer manual review. Names, codes and serial numbers deserve attention because dictionaries cannot reliably correct them.

Process a representative page first when the document is long. Pick one with ordinary paragraphs, a heading and a table or stamp. If that page performs poorly, fix rotation or image quality before spending time on the entire file. Large scans use significant memory in a browser; working in sections can be more reliable on a phone or older laptop.

Compare the recognized text with the page image. Check zero and the letter O, one and lowercase l, decimal separators, dates, currency values and reference numbers. Tables can lose columns because OCR reads text but must also infer layout. Handwriting, decorative fonts and low-contrast stamps are common failure points. If the transcript will drive a payment, legal filing or medical decision, have a person verify every critical value.

FeelPDF's OCR tool currently produces a text transcript rather than rebuilding a fully searchable visual PDF. That distinction is intentional and should be stated clearly. The transcript is convenient for search, notes and copying passages; the original scan remains the source for visual layout, signatures and seals.

Open the browser developer tools, select the Network panel and clear the existing entries after the OCR page and language model have loaded. Add the scan and start recognition. Model files or application code may have appeared during initial loading, but there should be no request whose payload contains your document. Running the test once gives you stronger evidence than relying on a privacy slogan.

Delete temporary downloads you no longer need, especially on a shared device. Give the transcript a name that identifies it as OCR output so nobody mistakes it for a certified transcription. Privacy is not only about the recognition request; it also includes where the source and result are stored after the task.