Make a scanned PDF searchable with local OCR

Why Ctrl+F does nothing in some PDFs

A PDF can look like a normal document while containing no usable text objects at all. Scanners often create one raster image per page. The page looks correct, but the PDF viewer has nothing textual to search.

OCR — optical character recognition — analyzes those page images and estimates the characters and words. A searchable-PDF workflow then places an invisible or visually unobtrusive text layer over the original scan so the document still looks like the source while search and copy begin to work.

Why local OCR matters

Receipts, contracts, invoices, school records, tax forms, identification documents, medical paperwork, and client drafts are exactly the files people may not want to upload to an unknown conversion site. Browser-side OCR changes the trust model: the document can be decoded, recognized, and rebuilt on the same device.

“Local” should mean more than a badge. A privacy-sensitive OCR tool should avoid sending the document bytes, page images, extracted text, or filenames to a remote processing service unless the user explicitly chooses a cloud feature.

Searchable PDF and text extraction are related but different

A good local OCR workflow

1. Open the scan locally. Confirm page count, orientation, and whether some pages already contain text.
2. Render only at the resolution you need. Extremely high DPI can waste memory without improving recognition.
3. Run OCR page by page. Language selection matters. Mixed Japanese/English, accented European text, numbers, tables, and handwriting have different error profiles.
4. Review low-confidence text. Names, totals, dates, serial numbers, and legal clauses deserve manual checking.
5. Export and test. Search for a phrase, copy a paragraph, and open the result in more than one PDF viewer if the document matters.

OCR does not magically understand the document

Recognition quality depends on scan resolution, blur, skew, contrast, font, language, handwriting, page curvature, tables, and compression artifacts. A clean 300-DPI printed page is a much easier problem than a phone photo of a folded receipt.

For important records, OCR should be treated as an assistive layer, not a replacement for the image. Keep the original scan so you can verify questionable characters later.

One synthetic page, measured on 25 October 2026

The OCRmyPDF README, opened 25 October 2026, says the program adds a searchable text layer with Tesseract, can keep the scan’s resolution, and produces PDF/A by default. It does not publish a byte size, a SHA-256, or a character count for a public fixture. MakeUseOf’s 15 October 2026 note (I made years of scanned PDFs searchable) says OCRmyPDF 17.13.0 found the word “award” on a private two-page scan where Acrobat’s search found nothing, and that one other word was misread. That scan is not in this repo. This machine has no ocrmypdf command. The walkthrough below is Tesseract 5.3.4, Pillow 10.2.0, and poppler’s pdftotext 24.02.0. Installed Tesseract language packs are eng, osd, por, and spa. The fixture is English. Japanese and other languages were not run.

Pillow drew a 1700×2200 white PNG in DejaVu Sans at 54 px. The four lines are Local OCR fixture 20261025. Invoice number 18439. Room 42B. Amount due is 19.50. Pillow then saved that image as a one-page PDF. pdftotext -layout on the image PDF returned only a form feed. pdffonts printed no font rows. pdfinfo reported PDF 1.4, one page, 816×1056 pts, title scan-clean, file size 85616 bytes.

pdftoppm -png -r 150 scan-clean.pdf took 0.137 seconds and wrote a 114675-byte PNG (SHA-256 2a8ed85f195c08e17f805cbbeac2a83a32c760593279008be2383adfbd286fca). tesseract scan-clean-ppm-1.png scan-clean-ocr -l eng txt pdf took 0.255 seconds. pdftotext -layout on that OCR PDF returns the four sentences with extra spaces between some words. After collapsing whitespace, the text matches the four lines. The Levenshtein distance on that collapsed string is 0.

The blurred page used the same render path. Tesseract took 0.243 seconds. Collapsing whitespace, the extract is Local OCR fixture 20261025. invoice number 18439. Room 426 Amount due is 19.50. Distance from the source sentence is 3: the capital I, the B in 42B, and the period after 42B. The layout extract also puts Amount on its own line. A heavier pass, width 500 and blur radius 2.2, made the text file empty (distance 80, the length of the source sentence). That image PDF is 68790 bytes, SHA-256 88792fdc68cab004d36ce186a5c417385b429d377ecc2bc40005698ee6169df3. None of these files were uploaded. A clean match on this one drawn page does not predict a phone photo of a folded receipt.

Why Ctrl+F fails, and what this run changed

Ctrl+F fails when the PDF has no text objects. This image PDF had none. Search started working only after Tesseract wrote GlyphLessFont into a second file. The first file was left as it was.

Does local OCR mean the text is exact?

Not on the blurred copy. Three edits in one short sentence is enough to change a room code from 42B to 426. Check names, totals, and serial numbers against the page image.

Did this run cover a real scanner or OCRmyPDF?

No. The page image was drawn by Pillow. OCRmyPDF was not installed, so deskew, --clean, and PDF/A were not measured. The September image benchmark below is still a separate test.

Engine tests are not a scanned-PDF walkthrough

The September 2026 OCR benchmark is an image-engine test on constructed pages. It is not a scanned-PDF UI, recognition, export, reopen, or search end-to-end. Fixture OCR-DE-2026-007 matched 11 of 11 key fields, but CER is 0.3023 and WER is 0.7381, so 100% fields is not perfect full-page recognition. Cross-fixture scores do not establish that folds or skew caused the other errors.

After you export a searchable PDF, reopen the file, search for a phrase you can see on the scan, and copy a paragraph in more than one viewer if the document matters. Treat the invisible text layer as a finding aid. Protocol: evaluation methodology.

What about existing PDF metadata?

Making a PDF searchable is not the same as sanitizing the whole PDF. A document can also contain author fields, creation software, XMP metadata, attachments, form data, annotations, JavaScript, signatures, and other structures. Some of those features can be damaged or invalidated by editing.

If privacy is the goal, inspect document metadata separately. The PDF metadata viewer reads author, creator, producer, title, and dates from an unlocked file in this browser. It does not read the words on the page, and it does not unlock an encrypted PDF. If authenticity is the goal, preserve signed originals and avoid pretending an edited PDF is still cryptographically identical.

When a searchable PDF is the right output

Use it when you want the visual fidelity of a scan plus the convenience of search: old manuals, paper archives, receipts, invoices, class notes, signed forms, research material, or books you are legally allowed to digitize. If you only need to copy a few lines from a screenshot, plain image-to-text OCR is faster.