You scanned a contract, or someone sent you a scanned report, and now Ctrl+F finds nothing. You can't select a sentence to copy it. To your eyes the page is full of text. To the computer it's a photo.
OCR (optical character recognition) fixes that. It reads the letters in the image and adds real text to the PDF, without changing how the pages look. Here's what's going on and how to get a clean result.
Why a scan is just a picture of text
A PDF made from Word or a web page stores actual characters: "the letter T, in this font, at this position". That's what search, copy and screen readers use.
A scanner or a phone scanning app doesn't know anything about letters. It captures pixels. The resulting PDF is one image per page, wrapped in a PDF container. Searching it is like searching a holiday photo for the word "beach". There's nothing there to match.
You can check in a few seconds: open the file and try to select a single word. If you can only draw a box around the whole page, or nothing highlights at all, it's an image-only scan.
What OCR adds to the file
OCR software looks at the image, finds lines and words, and works out which character each shape most likely is. Modern engines do this with a trained neural network for each language, which is why picking the right language matters so much.
The useful trick is what happens next. OCR PDF keeps the original page image exactly as it was and adds an invisible text layer, with each word placed where it appears in the image. The page looks identical. But now search finds words, you can select and copy text, and screen readers have something to read.
Because the visible page is still the original scan, an OCR mistake doesn't show up as a typo on the page. It only shows up when you search or copy.
Running OCR on EddiesPDF
- Open OCR PDF and add your scanned PDF.
- Set "Document language" to the main language of the text.
- If the document mixes two languages (an English contract with French annexes, say), set "Second language (optional)" as well. Don't add one you don't need, because it gives the engine more ways to misread a word.
- Under "Output format", choose PDF for a searchable copy, Text for a plain .txt file, or both.
- Under "Pages", leave "All pages" or enter a custom range like 1-3, 7 if you only need part of a long file.
- "Skip pages that already have text" is on by default (it's under More options). Leave it on for mixed files, so pages that already have real text aren't processed twice.
- Click Recognize text and wait for it to finish, then download.
The first time you use a language, its recognition data is downloaded from a public CDN. That's a one-off wait. Your PDF goes nowhere; the recognition itself happens in the browser.
That also means speed depends on your device. A 10-page document on a laptop is usually quick. A 200-page scan on an older phone can take a long while, and very large files may run out of memory. For long documents, a computer is the better choice, or process the file in chunks with the page range.
What makes OCR accurate (or not)
Most bad OCR results trace back to the scan itself.
Resolution
Around 300 dpi is the sweet spot for normal-sized printed text. Below about 200 dpi, small letters blur together and "rn" turns into "m". EddiesPDF renders each page at about 300 DPI before recognizing it, but that can't add detail the original scan doesn't have. If your scanner app has a quality setting, pick the higher one.
Straight pages
Text lines need to be roughly horizontal. A page photographed at an angle, or fed into a scanner crooked, gives noticeably worse results. Run the file through Deskew PDF first; it straightens tilted pages automatically. Cropping off dark borders and fingers with Crop PDF helps too.
Contrast
Black text on white paper works best. Faded thermal receipts, pale pencil, coloured backgrounds and heavy shadows from a phone camera all hurt. Scan in good, even light.
Printed text vs handwriting
OCR engines like the one we use are trained for printed text. Neat block capitals sometimes come out fine. Joined-up handwriting mostly doesn't, so don't expect a usable text layer from handwritten notes.
Layout matters less than people expect. Columns and tables are recognized as text, though the reading order in a plain .txt export can come out a bit jumbled on complex pages.
What to do with the result
A searchable PDF is often all you need: you can find clauses, copy a reference number and let your file search tool index the content.
If you need to edit the text, convert the OCR'd file with PDF to Word. That tool doesn't do OCR itself, so running OCR PDF first is required for scans. For pulling raw text into a spreadsheet or another program, the Text output here or PDF to text on the OCR'd PDF works well.
Check it by searching for a word
When the download finishes, open the PDF and search for a word you can see on page 1, then one from the last page. Pick something distinctive, like a name or a long word, rather than "the". If both are found, the text layer is there and it's lined up.
If a search fails on a word you can clearly see, try a few others. One miss is normal (OCR is never 100% accurate). If most searches fail, check the language setting first, then the scan quality. Those two account for nearly every bad result.
Also useful: Deskew PDF · PDF to Word · PDF to text
