OCR looks at a picture of a page and produces text. On a clean 300 DPI scan of ordinary printed text, accuracy is very high. On a phone photo taken at an angle in poor light, it is not.
The single biggest lever is scan quality, and it is decided before OCR runs. Rescanning at 300 DPI in greyscale beats any amount of software cleverness applied afterwards.
A scanned PDF contains no text. It contains pictures of pages, and the fact that those pictures show words is meaningful to you and to nothing else. Search returns nothing, copying is impossible, and a screen reader finds an empty document.
OCR is what closes that gap, and understanding roughly how it works explains both why it is so good on some documents and so poor on others.
What the software actually does
Recognition runs in stages, and each one can go wrong in its own way.
Layout analysis
Before reading anything, the engine works out the structure of the page: which regions are text, which are images, where the columns are and what order they should be read in. Mistakes here scramble the output even when every individual letter is read correctly.
Line and character segmentation
Text regions are split into lines, then into individual characters. Touching letters in a low resolution scan get merged into one shape, and broken letters in a faded one get split into two.
Classification
Each shape is compared against the engine's model of what letters look like and assigned the most likely candidate, with a confidence score.
Language correction
A dictionary and language model fix ambiguous cases. This is what turns rnodern into modern, and it is why choosing the right language matters so much.
rn from m, or l from 1, from shape alone. The language model resolves those cases by asking which reading produces a real word. Point it at the wrong language and that correction stage works against you, turning plausible text into confident nonsense.The four things that decide accuracy
Almost all of these are properties of the scan, not of the software.
- Resolution300 DPI is the target
- ContrastSharp black on white
- SkewStraight, under 1 degree
- TypefacePlain serif or sans
- Below 200 DPIAccuracy falls sharply
- HandwritingLargely unreadable
Resolution is the one to get right. Below about 200 DPI the individual strokes of a letter are only a couple of pixels wide, and characters begin to merge. At 300 DPI there is enough detail for the classifier to work with. Above 400 DPI there is very little further gain for printed text, and processing takes longer.
Enough detail to separate touching letters without inflating processing time.
Strokes become too thin to distinguish reliably. Rescanning beats post-processing.
Greyscale is a third the size and recognition is just as good, often better.
Two kinds of output
OCR can produce two quite different things, and choosing the wrong one is a common source of disappointment.
The page, plus hidden text
- Looks exactly like the original scan
- Invisible text layer sits behind the image
- Search and copy work
- Recognition errors are hidden but harmless
- Right for archiving and filing
Just the words
- Plain text with no images and no layout
- Much smaller file
- Every recognition error is visible
- Formatting, columns and tables are lost
- Right when you need to edit or reuse the words
What OCR still cannot do
Expectations have been raised by how well it handles clean printed text, and there are categories where it remains weak.
Handwriting is the main one. Cursive handwriting is effectively unreadable to standard OCR, and even neat printing gives poor results. Complex tables are another: the engine may read every cell correctly and still lose the structure that made the table meaningful. Multi-column layouts can be read in the wrong order if layout analysis misjudges the columns. And very stylised display typefaces, which differ substantially from the shapes the model learned, produce errors that the language model then confidently corrects into the wrong words.
Make a scan searchable in your browser
Recognition runs locally. The document is never uploaded and never leaves your machine.
Common questions
Getting a usable result.
01What does OCR actually do to my PDF?
It examines the picture of each page, works out where the text regions are, identifies each character shape, and writes the result as an invisible text layer positioned behind the original image. The page looks unchanged and becomes searchable.
02What scan resolution should I use?
300 DPI. Below about 200 the strokes of each letter become too thin to distinguish reliably and accuracy falls sharply. Above 400 there is very little further benefit for printed text and processing takes longer.
03Does scanning in colour improve accuracy?
No. Greyscale performs just as well and often slightly better, because recognition works on contrast rather than hue. Greyscale is also about a third the file size, which makes it the better choice on both counts.
04Why is my OCR output full of errors?
Usually the scan rather than the software. Low resolution, poor contrast, a skewed page or an unusual typeface all degrade recognition, and no amount of reprocessing recovers detail that was never captured. Rescanning at 300 DPI in greyscale, straight, fixes most cases.
05Can OCR read handwriting?
Standard OCR effectively cannot. Cursive is unreadable to it and even careful printing gives poor results. Specialised handwriting recognition exists and is a different technology with its own limits.
06Should I choose searchable PDF or text only?
Searchable PDF if you want to keep the document looking exactly as it does and simply be able to find things in it. Text only if you need the words themselves to edit or reuse, accepting that layout, columns and images are discarded.