A scanned contract is a photograph of a contract. You can read it, but your computer cannot: no search, no copy-and-paste, no way to find the indemnity clause across forty pages except by eye. OCR — optical character recognition — fixes that by reading the picture and adding the words as an invisible layer over the page.
The catch is that the documents people scan are exactly the ones they should not upload. Contracts, IDs, medical records, signed statements. Most OCR happens on someone else’s server. This guide does it on yours.
What OCR produces
Running OCR on a scanned PDF gives you two things:
- A searchable PDF. A new copy of the scan: each page is redrawn as a picture at up to 200 DPI, with an invisible text layer on top of it, aligned with the printed words. Search, selection and copy work as they would in a born-digital document.
- A plain-text file. The recognised words on their own, useful for pasting into a contract-review tool or a spreadsheet.
The page still looks like the scan, and the recognised words are only an addition over the picture, so a misread word changes what a search finds, not what the page shows. Keep your original scan alongside the searchable copy.
What decides accuracy
OCR is statistics, and its confidence rises and falls with the input:
- Resolution. 300 DPI is the working standard. A 150 DPI scan or a phone photo at arm’s length loses the small letters first — the serifs on an l and i, the difference between 5 and S.
- Contrast and cleanliness. Grey text on grey paper, coffee rings, and the shadow of a fold all cost characters.
- Skew. A page scanned at an angle confuses line detection. Rescan it flat; for a phone photo of a page, Scan PDF straightens it.
- Typeface. Clean printed text recognises well. Stylised fonts less so. Handwriting, honestly, is a gamble.
- Language. Filozy’s engine recognises English. Other languages will come out garbled.
You cannot fix a bad scan with a better algorithm. Rescanning at 300 DPI, flat, in good light, is worth more than any setting.
Step by step
- Trim first. Auto crop trims plain margins around the text. It does not straighten pages, and it leaves alone a page whose border is not one plain colour, such as a dark scanner edge, so crop that one by hand in Crop PDF. Clean edges help the engine find the text block.
- Open OCR PDF and choose the trimmed file. The recognition engine is downloaded from Filozy’s own servers at this point — it is program code, not your document going the other way.
- Run the tool. A progress bar counts the pages. At the end the result gives one average confidence figure for the whole run; below about 80, read the text through before you rely on it.
- Save the searchable PDF, and the text file if you want it.
- Verify before you rely on it. Search the result for a clause you know is there. Compare every number, date, party name and defined term you will act on against the page image. OCR errors cluster in exactly those places: 0 and O, 1 and l, decimal points, currency symbols.
What to do next
The searchable PDF is the input to everything else:
- Need to remove personal information before sharing? Redact PDF works from boxes you draw. It has no search and runs no OCR, so mark every occurrence yourself; searching this copy in your PDF reader first shows where each one is.
- Filing or producing the document? Add Bates numbers with Header, footer & Bates.
- Sending it on? Protect PDF encrypts it on the same device.
A ready-made sequence for this — trim, recognise, number, protect — exists as the Prepare a scanned contract for filing workflow, which carries the file from one tool to the next without re-selecting it.
Honest limits
- Tables are recognised as text, not as cells. For a spreadsheet, run the result through PDF to Excel and expect to check the columns.
- Handwritten annotations are usually missed or misread. If a handwritten date matters, read it yourself.
- Very long documents take time, and a phone has less memory than a laptop. Split a 400-page scan into parts first.
- The text layer is only as good as the scan. A confidence figure of 90% on a legal document means one word in ten may be wrong. That is why step 5 is not optional.