How to Make a Scanned PDF Searchable Without Uploading It
A scan looks like a document and behaves like a photograph. You can read it. Ctrl+F cannot. OCR guesses the characters and writes them as a text layer so search and copy work. The guess is only as good as the print, the resolution, and the language. On this site the language is English, and the guess is not a proofreading pass.
A scan is a picture until something reads it
A flatbed or a phone camera stores a PDF page as an image of paper. You can see the words. The computer cannot select them, search them, or copy them, because they are pixels, not text. Optical character recognition looks at those pixels, guesses the characters, and writes the guesses back as a text layer. The picture stays. The words become selectable.
The PDF Suite does this with Tesseract.js 5.1.0, in the browser, when you check OCR before you save. The worker is created for English (eng). The first run downloads the English model. The page image is then recognized in a worker in the tab. It is not sent to a cloud OCR API by this code. What else the page loads, and what it does not, is in What Leaves Your Browser.
OCR is not a compressor. A searchable scan can be as large as the scan you started with, or larger, because you added a text layer. If the problem is a portal's file-size cap, read How to Shrink a PDF Locally and pick a preset. If the problem is "I need to find the clause," this is the right page. You can do both, in that order: keep a readable bitmap, then recognize it. Crushing a scan to the Maximum preset and then expecting perfect OCR is how you get wrong words.
What the suite does on each page
Pages are not all treated the same. Before OCR, the suite asks pdf.js for the page's text items. If any text is already there, that page is copied as a normal PDF page and OCR is skipped. A digital contract with a real text layer stays a digital contract. A scan with no text layer is rendered and passed to Tesseract.
The render used for recognition is the same high-resolution render the image exporter uses: the pdf.js viewport at scale 2, encoded as JPEG at quality 0.9. Tesseract is asked for a single-page PDF (its pdf: true output). That page, which contains the recognized text plus Tesseract's image, replaces the original page in the file you download. The original vector page is not kept underneath for pages that went through OCR, because those pages had no text layer to keep.
Rotation is applied in the suite before export. Fix sideways pages before you recognize them. Tesseract will try to read a page that is on its side, and it will do a bad job. Delete blank pages first so you do not spend the recognition time on them. The page grid is where you rotate, reorder, and delete. Undo exists, including Ctrl+Z or Cmd+Z.
After the document is assembled, the compression preset runs on the whole file if you did not leave it on No Compression. A heavy preset can rasterize pages you just recognized and can damage the text layer you just added. For a file you mainly want to search, start with No Compression or Lossless, confirm search works, and only then try Balanced if you still need the bytes. Check search again after the preset.
English only, and the other limits
- The model is English. There is no language menu. French, German, and Spanish with mostly Latin letters often produce partial results and wrong accents. Chinese, Japanese, Korean, Arabic, and Hebrew will not come out right. Do not use this checkbox for those documents and then ship the text layer as if it were a translation.
- Handwriting is out of scope. Tesseract is built for printed type. A signature, a doctor's note, and a form filled in pen will be garbage or empty. Leave them as images.
- Small type and low contrast fail quietly. Footnotes, stamps, and gray text on a gray scan produce confident wrong words. The engine does not raise its hand. You have to look.
- Layout is approximate. Multi-column pages, tables, and sidebars can be read in an order a person would not use. Search still finds a word. Copy-paste of a whole page may scramble the columns.
- The first run needs a network to fetch the engine and the English model. Later runs in the same session can work from what the browser cached. If the fetch fails, the suite alerts you and falls back to a standard export without OCR. You do not get a half-labeled success.
A scan at roughly 300 DPI of printed body text is the case this is good at. The scanner advice in the shrink guide (do not scan text at 600 DPI color "just in case") still holds. More pixels help only until the letters are clearly resolved. After that they cost time and memory. Very small scans, under about 150 DPI of 10-point type, are fuzzy to the engine for the same reason they are fuzzy to you.
Steps
- Open the PDF Suite and drop the scan. Wait until every page has a thumbnail. If a page is sideways, rotate it on the card.
- Delete blanks and covers you do not need. Recognition time is per page that lacks a text layer.
- Set the compression preset to No Compression or Lossless for the first export. Check OCR.
- Save. The status line shows recognition progress. A long scan takes a while because it is your CPU, not a queue on a server. Leave the tab in the foreground if the browser starts throttling background tabs.
- Open the download. Press Ctrl+F or Cmd+F and search for a distinctive word you can see on the page, not a word from the header that might have come from somewhere else. Select a line in the middle of a paragraph and copy it into a text editor.
- Read that line against the paper. If characters are substituted (rn for m, 0 for O, dropped accents), the layer is a finding aid, not a quotation source. Do not paste it into a filing as if it were the record. Quote from the image, or correct the text.
- If you still have to hit a size cap, make a second export with Balanced, from the original scan, and run the same search. If search dies, the preset rebuilt the pages as images. Stop and remove pages or lower the scan resolution instead. The preset behavior is described in Lossy vs Lossless.
Photographs of paper, and pages that already have text
A phone photo of a letter is a JPEG, not a PDF. Drop it on the PDF Suite anyway. The suite accepts images and builds a page from them. Check OCR if you need the words. Straighten the photo first in Image Studio if the page is skewed. Tesseract assumes lines of type are roughly horizontal. A strong keystone from a phone held at an angle will cost you accuracy, and neither tool has a perspective-correction control. Retake the photo square if the text matters.
A PDF that is already digital does not need OCR. The suite will notice the text and skip the page. If you check the box anyway, you pay the cost of the text-detection pass and you keep the original page. You do not get a second, conflicting layer from this path. If search already works, leave the box unchecked and save yourself the model download.
Password-protected files have to be unlocked before the suite can read them. The shrink guide's note on that still applies. OCR cannot see through encryption.
What to do with a bad layer
Do not "fix" a bad recognition by running OCR again on the OCR output. You would be recognizing a JPEG of an already-recognized page. Go back to the scan. Improve the input: rotate, crop the margin in an image tool if the page is a photo with a table in the background, or export a cleaner render. Then recognize once.
If the document is not English, use a desktop OCR engine that lets you pick the language, on a machine you are willing to open the file on. This checkbox will not grow a language menu by being run twice. If the document is a contract and the text layer will be copied into an email, have a person read the copied sentence against the scan. Recognition errors in names and amounts are the expensive kind.
Hidden metadata is a separate pass. A freshly built file from this suite drops a lot of the source document's old structure because pages are rewritten, and it can still carry a watermark you typed yourself. Author dictionaries and XMP are discussed in Remove Hidden PDF Metadata. OCR does not redact. It adds text. If the scan shows a social security number, the text layer now contains that number in a form that is easier to copy. That is useful for you and it is also easier to leak. Treat the searchable file as more portable, not as more private.
Frequently Asked Questions
Which languages can the OCR checkbox read?
English. The worker is created with the eng model, and there is no language picker. Other Latin-script languages may be partly recognized. CJK, Arabic, and Hebrew will not come out right. Handwriting is out of scope.
Will OCR shrink the PDF?
No. It adds a text layer. Shrink the file with a preset only after you have confirmed search, and confirm search again, because a heavy preset can rebuild pages as images.
Does a digital PDF need OCR?
No. If the page already has text items, the suite skips recognition and copies the page. If search already works, leave the box unchecked.
The words look almost right. Can I paste them into an email?
Only after you compare the sentence to the scan. Names, amounts, and dates are where a one-character error is expensive. Treat a messy layer as a way to find the page, not as the record.
Add a text layer to a scan
Rotate the pages first, leave compression on Lossless or off for the first export, and check the words.
Open the PDF Suite