Image-based vs text-based PDFs: how to tell, and why it matters
Two PDFs can show the same page and be nothing alike inside. One holds real characters with fonts and positions; the other holds a photograph of those characters. Everything that frustrates people about PDFs — search that finds nothing, copy that produces gibberish, huge files, useless screen readers — comes down to that difference.
The five-second test
- Try to select a line of text with the mouse. A text-based PDF highlights words; an image-based one draws a rectangle across the whole page, or nothing at all.
- Press Ctrl+F and search for a word you can clearly see. No match means no text layer.
- Zoom to 400%. Real text stays crisp at any zoom; an image gets soft or blocky.
- Check the size. Twenty pages of text in 200 KB is a text-based file; twenty pages in 25 MB is scanned images.
A third case exists and is the best of both: a scanned page with an invisible text layer added afterwards. It looks like an image, selects like text, and is what a good scanning workflow produces.
Where each kind comes from
Text-based PDFs are exported from software — a word processor, a layout program, a spreadsheet, a report generator. The file records which glyph sits where, in which font, and often keeps structure such as headings, tables and links.
Image-based PDFs come from cameras and scanners, from screenshots, and from any conversion that starts with pictures — including a set of flipbook page images. The file records pixels, and the meaning of those pixels is left to the human eye.
What the difference costs you
- Search and citation: impossible without a text layer, which makes long documents much harder to work with.
- Accessibility: screen readers and braille displays need text. An image-only PDF is effectively blank to them, and in many places that is a legal problem for a public document.
- Translation and summarising: any tool that reads a document needs characters, not pixels.
- Reflow on small screens: phones can re-wrap real text to the screen width; an image can only be zoomed and panned.
- File size and battery: images are heavier to store, to send and to render.
- Editing: correcting a typo in a text-based PDF is fiddly but possible; in an image it means re-scanning or retouching pixels.
Image-based files do have genuine strengths. They preserve exactly what the original looked like, including handwriting, stamps, signatures and paper damage, which is why archives and legal evidence often prefer them. They also render identically everywhere, with no font substitution surprises.
What OCR actually does
Optical character recognition looks at the image, recognises shapes as characters, and writes an invisible text layer positioned over the picture. The visible page does not change at all; you simply gain selection, search and screen-reader support.
Accuracy depends on the input. Clean 300 DPI scans of ordinary print reach the high nineties per cent. Accuracy drops with low resolution, skew, shadows, tight columns, tables, handwriting, unusual typefaces, and any language whose model you have not installed. Numbers and proper nouns are the usual casualties, which matters if you plan to cite figures.
Practical habits that help: scan or export at 300 DPI, keep pages straight, use greyscale rather than colour for plain text, and tell the OCR tool which languages to expect. Always spot-check a few pages by searching for a word you know is there.
Choosing deliberately
If you control the export, always produce a text-based PDF: it is smaller, searchable and accessible, and it can still be printed perfectly. If you are working from pictures — scans, photographs or page images — accept that you have an image-based file, then decide whether the document deserves an OCR pass.
The rule of thumb: if you will read it once, leave it alone. If you will search it, quote it, share it with colleagues or keep it for years, run OCR and store the result. Five minutes of processing turns a stack of pictures into a document you can actually use, and it costs nothing in visual fidelity.