OCR
OCR a PDF: Make Scanned Documents Searchable and Copyable
A scan or a photo saved as a PDF is really just an image. You can look at it, but you cannot select the text, search it or copy it — after a hundred archived pages, finding one sentence means scrolling.
Why a scan cannot be searched
In a normal PDF every character carries a text layer: what it is and where it sits. Search, selection and copy all work off that layer.
A scan from a scanner or a phone has no such layer — it is a single bitmap. The reader sees a picture, so full-text search returns nothing, dragging selects a blank area, and copying produces nothing. The file is not corrupt; it simply never contained text.
How OCR turns the image back into text
OCR (optical character recognition) looks at pixel shapes and decides which character they represent. UnipagePDF does this in four steps:
- Render each page as a bitmap at 2× resolution, so letter edges stay crisp
- Run recognition page by page, recording the position of every word
- Build an invisible text layer aligned to the original glyphs
- Assemble a new PDF that looks identical but is selectable, searchable and copyable
The three settings
| Setting | What it does |
|---|---|
| Language | English, Chinese, or mixed English+Chinese. Picking the wrong one noticeably lowers accuracy; mixed documents should use the mixed option |
| Page range | Defaults to every page. Enter 1-5 or 1,3,5 to process only what you need |
| Progress bar | Shows page-by-page progress so you can see where a long job has got to |
The first run loads the recognition engine
The engine and its language data run locally in your browser — your file is never uploaded — but the first use downloads the language pack (about 11 MB for English, 20 MB for Chinese). Until that finishes, the status sits at "Loading OCR engine". Once cached, later runs start much faster.
If that step seems slow, restrict the job to a page or two with the page range field, check the result, then process the whole file.
What the result can and cannot do
- It is not 100% accurate: clean printed body text usually works well, while handwriting, highlighter marks, heavy skew and complex layouts frequently fail.
- Small or low-contrast pages lose words more often — spot-check the search results on a few pages after exporting.
- The text layer is added on top: it aligns to the original image, which stays underneath. If the source image is blurry, the recognised text inherits that.
Order matters when combining tools
Two combinations come up constantly, and the order changes the outcome:
| What you want | Correct order |
|---|---|
| Find IDs and phone numbers with Redact's batch find | OCR first, then redact — a scan with no text layer cannot be searched |
| Rename files by their content | OCR first, then rename — renaming also needs the text layer |
| Recognise and then shrink the file | OCR first to create the text layer, compress last |
OCR adds a text layer and deletes nothing. To remove metadata, comments or form data afterwards, use Clean Private Data.
Is the file uploaded?
No. The engine, the language data and the whole pipeline run in your browser; the file never leaves your device.