Scanned PDF or text PDF? Check first

PDFs come in two kinds, and it matters which one you have. A PDF exported from a word processor contains real text: the characters themselves, positioned on the page. A PDF made by a scanner or a phone camera usually contains only a picture of each page.

The quickest test: try to select a word. If a highlight snaps to the words, the page has a text layer and doesn't need OCR. Reading that text directly is exact, while OCR is always an estimate. If dragging draws a box, or selects nothing, the page is an image and OCR is what you need.

Many documents are mixed: a typed contract with a scanned signature page, or a report with a scanned appendix. Vault Studio Pro checks each page separately, reads the text layer where there is one, and only runs OCR on pages that are pictures.

How OCR works

Modern OCR engines, including Tesseract, the open-source engine Vault Studio Pro uses, work in stages:

  1. Clean-up. The image is converted to black and white (binarised), so the engine can separate ink from paper. Uneven lighting, shadows and coloured backgrounds make this step harder.
  2. Layout analysis. The engine finds blocks of text, then lines within them, and works out their reading order. Columns, tables and text wrapped around pictures are where this step can go wrong.
  3. Line recognition. Since version 4, Tesseract recognises whole lines with a neural network (an LSTM), rather than matching one character shape at a time. Reading a line in context helps with touching or broken letters.
  4. Language data. A trained model for each language knows its alphabet and common letter sequences. That's why OCR needs to know the language of the page, and why an English model struggles with other scripts.

The output is text, a position for each word on the page, and a confidence score for each word, which is the engine's own estimate of how sure it is.

Running OCR in the browser

Most online OCR services upload your scan and run the engine on their servers. That's often the very document you'd least like to upload: an ID, a medical letter, a signed contract, a bank statement.

Tesseract.js, an open-source project, compiles the Tesseract engine to WebAssembly so it can run inside a web page. Vault Studio Pro uses it like this:

  • The OCR engine and its English language data are served from our own domain, not a public CDN, and the page's security policy blocks connections to other sites.
  • Recognition runs in a web worker, a background thread of the page, so the editor stays responsive while a page is being read.
  • Each scanned page is rendered in the page and passed to the engine as an image. The scan never leaves your device.

If you want OCR available without a connection, choose "Make available offline" in the app's Privacy Audit panel first. It stores the engine and language data in your browser. Our guide to what "client-side" really means shows how to confirm that nothing is uploaded.

Step by step: OCR a scanned PDF in Vault Studio Pro

  1. Open the PDF editor and open your scan, or drop it onto the window.
  2. Go to the page you need and press OCR Page. The OCR panel opens and shows whether the text came from the PDF's own text layer (exact) or from OCR, with the engine's confidence and how long it took.
  3. Read the recognised text in the panel and correct anything important. Copy puts it on your clipboard.
  4. With Pro, use OCR All to read every page, then Send to Vault Docs to get an editable .docx, or Extract Tables to Sheets to turn tabular text into rows and columns.

One thing to know: the recognised text lives in the app. Saving the PDF doesn't add an invisible searchable text layer to the file, so to keep the text, copy it or send it to Docs.

What is free and what needs Pro

  • Free: reading any page that already has a text layer, and OCR on the first page of each PDF. That covers one-page letters, receipts and forms, and lets you see how well OCR reads your documents before paying anything.
  • Pro (ocr on every page): OCR beyond the first page, OCR All, and sending recognised text to Docs or Sheets.

Every Pro feature is unlocked for the first 7 days after you first open the app, with no card required. After that, a one-off Week Pass covers a single job, and monthly, yearly and lifetime plans suit regular use. See pricing for the current prices.

Tips for better accuracy

Most OCR errors come from the image, not the engine. Tesseract's own documentation is a good guide, and these points follow from it:

  • Scan at 300 dpi. Tesseract's documentation says it works best on images of at least 300 dpi. Much lower and letters blur together. Much higher mostly makes files bigger.
  • Keep the text large enough. The same documentation notes the best results when capital letters are roughly 20 to 40 pixels tall. Tiny footnotes and fine print are the most error-prone. The app renders each scanned page at twice its PDF page size before recognition, so a low-resolution scan can't be rescued by the renderer.
  • Straighten and flatten. Skewed or curved lines, like a phone photo of an open book, confuse line finding. Scan flat, or photograph straight on in even light.
  • Boost contrast on photos. For a phone photo of a page, Image Studio has a black-and-white "document scan" preset, plus crop and rotate. Clean the image, then OCR it.
  • Crop away clutter. Hole punches, dark scanner edges and fingers can be read as characters.
  • Proofread numbers. Confusions like 0 and O, 1 and l, or 5 and S matter most in amounts, dates and account numbers. Check the confidence score and read those fields against the original.

Checking the result

OCR output is a draft, not a copy. A sensible routine for anything that matters:

  1. Look at the confidence the panel reports. A low score on a page is a cue to read that page closely.
  2. Compare every number, name and date against the scan, side by side. These are the fields where a single wrong character changes the meaning, and where a spell checker won't help.
  3. Check the reading order on pages with columns, sidebars or tables.
  4. Keep the original scan as the record. If the text will be relied on later, note that it was produced by OCR and checked by hand.

Limits worth knowing

  • English only. The app ships the English language model. Other languages, and especially other scripts, will be recognised poorly or not at all.
  • Handwriting is mostly out of reach for this kind of engine. Printed text works best.
  • Complex layouts, such as multi-column newsletters, forms with boxes and tables without ruling lines, may come out in the wrong order. Table extraction works best on simple, well-aligned tables.
  • Large documents take time, since each page is processed on your own computer, and the speed depends on that computer.
  • OCR and redaction: redacted pages are saved as images, so their remaining text isn't selectable. You can OCR them afterwards to read the rest of the page. See how to redact a PDF properly.

Try it

Open a scan in the PDF editor and press OCR Page. The PDF editor page lists everything else it does, and Image Studio can read text from a single photo or screenshot the same way.

OCR a PDF nowSee plans and pricing

Sources

  1. Improving the quality of the output. Tesseract documentation
  2. Tesseract Open Source OCR Engine. project README on GitHub
  3. Tesseract.js: the Tesseract engine as a WebAssembly port for browsers and Node.js. project README on GitHub
  4. Web Workers API. MDN Web Docs