PDF Text Recognition: What OCR Is and What It Gives You
By Schema LawPublished Updated
PDF text recognition is the step that turns a picture of a page into characters a computer can search, copy and export. It is the same thing as OCR, which stands for optical character recognition. It gives you text that follows the page closely, and it leaves the job of checking that text to you.
If you are new to the subject, start here. The sections below say what recognition does to a scan, sort out the names it goes by, list what the software needs and what it hands back, and set out how much weight the result can bear in legal work.
What does text recognition do to a scanned PDF?
Text recognition looks at the picture on each page of a scanned PDF, decides which marks are letters, digits and punctuation, and records them as characters. Before recognition the file holds only dots. Afterwards there is a set of characters that says what the dots appear to spell.
A scanner does not know what a letter is. It measures how dark the paper is at several million points and stores those measurements. To you the result is a promissory note. To the computer it is a grid of light and dark, no different in kind from a photograph of a parking lot. The post on why you cannot copy text from a PDF describes that kind of file under what an image-only PDF is.
Take one line from a scanned note between two fictional parties:
The Borrower, Calloway Marine Supply LLC, promises to pay $42,750.00 on or before June 30, 2023.
Before recognition, a search for "Calloway" finds nothing, because the file contains no letter C followed by a letter a. After recognition, the computer has a string of about ninety characters for that line. Search finds the name. You can copy the sentence into a demand letter. A spreadsheet can receive the amount as a number.
Use the new characters to search. On a deed, look up each party and the recording citation. On a medical bill, look up the total. You decide which line states an obligation by reading the page next to the words. Recognition has handed you the characters and stopped there.
Here is the same change in terms of what you can do with the file.
| Task | Scanned PDF before recognition | After recognition |
|---|---|---|
| Read it on screen | Yes | Yes |
| Print it | Yes | Yes |
| Search for a name | No | Yes, where the name was read correctly |
| Copy a clause | No | Yes |
| Export to Word or a spreadsheet | No | Yes |
| Have it indexed by a document system | By file name only | By its contents |
Two things do not change. The paper original is untouched, and so is the picture of it, unless you ask the software to alter the image. Recognition adds a description of the page. It does not replace the page.
Where the new characters are kept depends on the tool and the export you choose. They may be tucked into the same PDF, or written to a separate document. The post on what an OCR file is walks through each of those containers.
Is PDF text recognition the same thing as OCR?
Yes. PDF text recognition, PDF character recognition and PDF optical character recognition are three names for one process, and OCR is the abbreviation of the third. Software menus, scanner panels and vendor pages each pick a different name, which makes one feature look like several.
The spread of names is an accident of history and marketing. "Optical" is there because the input is something seen, a picture, as opposed to keystrokes. "Character" is there because the early machines worked one letter at a time. "Text recognition" is the plain-language version that menu designers prefer.
| Name you may see | Where it tends to appear | Same thing as OCR? |
|---|---|---|
| Optical character recognition, OCR | Technical writing, court rules, vendor pages | Yes |
| Character recognition | Older manuals, scanner drivers | Yes |
| Text recognition, recognize text | Menus in PDF programs and phone apps | Yes |
| Scan text recognition, scanned document text recognition | Search queries and help pages | Yes, said with the input named |
| Make searchable | Copier panels, PDF programs | Yes, with a searchable PDF as the output |
| Text extraction | Review platforms, developer tools | Not always; see below |
The one neighbor that is different
"Text extraction" sometimes means recognition and sometimes means something else. When a PDF was saved from a word processor, its characters are already in the file, and extracting them is copying. No picture is read and no guess is made. When someone tells you text was "extracted" from a scan, they mean it was recognized. If the distinction matters to your task, ask which it was.
Handwriting and marks
Recognition of hand-printed characters and of checkboxes goes by other abbreviations, and vendors use them loosely. You do not need them to follow this post. The OCR glossary for lawyers gives a one-line definition of each term in the family and points to the post that covers it.
What goes into text recognition and what comes out
A page image goes in and characters come out. The image can be a page of a scanned PDF, a fax, a PNG or JPEG, or a phone photo. The output is always text, and depending on the tool it may also include where each word sat and what structure the page had.
What goes in
| Input | What the software can use | What limits it |
|---|---|---|
| The page picture | The shapes of the marks | Faint print, small type, specks, a tilted page |
| Resolution and color | Detail in each letter | A coarse or heavily compressed scan |
| The page as a whole | Columns, tables, headings | Crowded layouts, stamps over text |
| Sometimes, a language setting | Likely words and alphabets | A wrong setting, or mixed languages |
Nothing else goes in. The software has no access to the paper, to the person who typed it, or to your knowledge of the matter. Everything it tells you was worked out from the picture.
What comes out
- Characters. Always. This is the reading itself.
- An order. The characters are in a sequence, which is the software's view of how the page should be read.
- Positions. Some tools record where on the page each word or line was found.
- Structure. Some tools mark which text was a heading, a list or a table cell.
- A measure of doubt. Some tools report how sure they were. Many do not.
What you receive depends on the product. Schema Law, for instance, accepts a PDF, PNG or JPEG of up to 20 MB, reads each page into text and tables, and shows you the result page by page before you export it. The how it works section of the homepage describes those three steps.
How the software gets from picture to characters is a subject of its own. Older engines work in fixed stages, and newer ones read the page with a trained model. The post on AI OCR explains how classic OCR works, stage by stage.
What text recognition does not do for a legal document
Text recognition does not understand a legal document, does not verify it and does not correct it. It reports which characters the page seems to show. Whether those characters form a valid notice, an authentic signature block or the right dollar figure is outside what it does.
Each of the three limits has consequences that are easy to miss.
It does not understand. The software can read "This Agreement shall terminate on March 1, 2025" without knowing that it has found a termination date. It treats that line the same way it treats the fax header. Finding the governing-law clause or the notice period is a separate step that works on the recognized text, and it inherits every error in that text.
It does not verify. A page that was altered before it was scanned is read as faithfully as a page that was not. Recognition tells you nothing about who signed, whether a page was swapped, or whether the copy is complete. If page 7 of a twelve-page lease is missing from the scan, you get eleven pages of text and no warning.
It does not correct. If the original says "Febuary", a faithful reading says "Febuary". That is what you want, because the page is the record. The reverse also happens: some tools lean on a dictionary or a language model, and may hand back "February" where the page has the misspelling. Either way, the recognized text is not a proofread version of the document.
| What you might assume | What is actually true |
|---|---|
| The text is the document | The text is one reading of a picture of the document |
| A clean-looking paragraph was read correctly | A wrong digit looks as tidy as a right one |
| Every page was read | A blank or failed page produces no text and may produce no notice |
| Handwritten notes were captured | They may be skipped or guessed |
| Redacted words are gone | Recognition does not remove anything from a file |
None of this is a reason to avoid the technology. It is a reason to be exact about what you have after using it: searchable working text, with the scan still serving as the document.
Where a law office meets text recognition without asking for it
A law office meets text recognition in at least four places it never chose: the copier, the PDF program, the document management system and the phone. Each can add recognized text to a file quietly, so many scans in your folders carry text whose origin nobody remembers.
- The copier or scanner. A "searchable PDF" option in the scan settings puts text behind the pages of every scan once someone turns it on.
- The PDF program. Desktop PDF software can recognize text as part of another command, such as editing or exporting a scan.
- The document management or review system. A system that indexes files by content can run recognition on files without text. The text may live in the index and never appear in the file you download.
- The phone. A scanning app or phone camera that lets you select words in a photo is running recognition.
How each of these behaves, and what to find out about a step that runs by default, is covered under what auto OCR is and where it runs unasked.
Why does this matter? Because the quality of that silent text varies with whatever produced it, and the file does not say which program that was. Two habits help.
- Do not assume a file with text was typed. A scan with a copier-made text layer passes a quick search test and can still be full of misreadings.
- Do not assume a file without text was never recognized. The system that indexed it may hold text you cannot see or export.
When a search result is going to carry weight, find out where the text came from. If you cannot, run the pages through a tool you have tested and compare.
Can you rely on recognized text without reading the page?
You can rely on recognized text to find things, and you cannot rely on it to prove things. Use it to locate pages, build a first draft and narrow a review. Before a name, date, amount or quotation leaves your office, read it against the page picture.
How much checking a task needs depends on what an error would cost.
| Use of the text | Reliance without reading | What to check |
|---|---|---|
| Finding which pages mention a party | Reasonable, with several spellings | A sample of pages the search did not return |
| Getting the gist of a long record | Reasonable | Nothing, until you cite it |
| Drafting from a scanned agreement | As a starting draft only | Every defined term, number and cross-reference |
| Quoting in a brief or letter | No | The full quotation, character by character |
| Totals from a scanned ledger | No | Each figure, then the arithmetic |
| Stating that a document does not contain a term | No | The pages themselves, or a second method |
The last row deserves a comment. A search that finds a word tells you something definite. A search that finds nothing tells you only that the word is not in the recognized text in the form you typed. Those are different facts.
A short routine for any recognized document
- Count the pages in the scan and the pages that produced text. They should match.
- Pick the worst-looking page and read its text against the picture.
- Check every proper name once, at its first appearance.
- Check every number you intend to use.
- Note in the file what you checked and what you did not.
That routine is the short form. The full method for proofreading OCR output against the original is in the post on why OCR output is full of errors.
Schema Law's own answer on this point is blunt. Nothing in its output marks a passage as uncertain, so the homepage's questions section tells you to check names, dates and amounts against the original. The same advice applies to any tool, including ones that do report doubt, because a confident misreading is still a misreading. Treat this as document-handling practice, not legal advice; what your court or client requires of a quoted or produced document is for you to confirm.