How to tell a PDF needs OCR
A PDF from a scanner, a phone scanning app or a fax gateway usually contains one picture per page and nothing else. It looks like a document, but to the computer it is a photograph. The signs are easy to spot:
- Ctrl+F (Cmd+F on a Mac) finds nothing, even for a word you can clearly see.
- Dragging the mouse selects the whole page as a block instead of individual words.
- Copying and pasting produces nothing, or the reader offers to "recognise text".
- Document search in your email, cloud drive or archive never finds the file.
OCR fixes all four at once. Typical candidates are signed contracts that went through a scanner, old paper archives, receipts and invoices for bookkeeping, and book chapters scanned for study.
What the server does with your file
The conversion runs OCRmyPDF with the Tesseract recognition engine on our server. For each page, it renders the page image, recognises the words in the chosen language, and adds them back as an invisible text layer placed precisely over the matching words in the picture. The image itself is not redrawn or altered, which is why the result looks exactly like the original.
A few details are worth knowing:
- Pages that already contain text are skipped, so a document mixing typed and scanned pages keeps its original text untouched.
- A light, lossless optimisation pass is applied, so the file size usually stays close to the original.
- The output is saved as PDF/A, the archival variant of PDF designed to display the same way decades from now. Any normal PDF reader opens it.
- One language is used per run. For a document that is mostly in one language with a few foreign names, choose the main language.
Getting accurate text
Recognition quality is decided mostly before the file reaches the server.
Resolution
Aim for 300 dpi when scanning printed documents. At that resolution normal 10 to 12 point text is easily recognised. 200 dpi works for clean, larger print. At 150 dpi or below, small letters merge and errors such as "rn" read as "m" become frequent. Scanning higher than 400 dpi rarely improves results and makes the file much bigger.
Orientation and straightness
This tool does not rotate or straighten pages. A page scanned sideways or upside down gives poor or empty text. Fix the orientation first with Rotate PDF. A slight tilt of a degree or two is tolerated, but a phone photo taken at an angle, with the page curving near the spine of a book, loses accuracy quickly.
Contrast and clutter
Black text on white paper is ideal. Faded thermal receipts, pencil, coloured paper, heavy stamps over the text and background patterns all reduce accuracy. Handwriting is not recognised in a useful way.
Long documents and time limits
Recognition takes a few seconds per page, so a 50-page scan typically finishes in a few minutes. Jobs have a processing time limit of 15 minutes without an account and 60 minutes on Premium. A book of several hundred pages may not fit into the free window; split it into parts with Split PDF and process them one by one. The daily allowance is one free run for files up to 500 MB, or unlimited runs up to 4 GB on Premium.
Password-protected PDFs are refused, because the pages cannot be read without the password. Remove it first if you have it. The uploaded file is deleted when processing ends, and the result expires about an hour after the job was created.
Checking the result
Open the searchable PDF and search for a few words from different pages, including a number or two. OCR errors cluster in predictable places:
- Look-alike characters: 0 and O, 1, l and I, 5 and S, especially in codes and account numbers.
- Tables, which come out as lines of text rather than cells.
- Small print in footers and stamps.
For contracts or invoices where an amount matters, always read the figure from the image, not from copied text. If you want the text itself rather than a searchable PDF, open the result in PDF to text and it will extract the recognised layer.