Extracting Text from Scanned PDFs: How OCR Works and Where It Fails
You open a PDF, try to select a sentence, and nothing highlights. Or you search for a name you can clearly see and the viewer finds nothing. That PDF is almost certainly a scan: a picture of a page rather than a page of text. Getting the words out of it requires optical character recognition, usually shortened to OCR. This guide explains how to confirm what you have, what OCR really does, why its results are never perfect, and how to improve your odds. It also tells you where our own tools stand today.
Scanned PDF versus text PDF
A PDF created from a word processor, a web page or an export contains real text with fonts. You can select it, copy it and search it. A scanned PDF, or a PDF made from photographs, contains an image of each page. To a computer, the letters are patterns of pixels, no different from a picture of a landscape. Some scanners add a hidden text layer after recognising the words, which makes a scan searchable while still looking like a scan; others do not. A quick test: try selecting a word, or use the search function. If neither works, there is no usable text layer.
Test it with PDF to Text
PDF to Text copies out the text that a PDF actually contains, and it works entirely on your device. If you run a scanned PDF through it, the tool reports that no text was found, and tells you the file looks like a scan. That is a useful diagnosis in itself: it tells you whether you need OCR or whether the text is already there and only needs extracting. If the tool returns text, you may not need OCR at all.
The same reasoning explains failures elsewhere. Converting a scanned PDF to Word with PDF to Word works only with PDFs that already contain real text. Scans have none to recover.
What OCR does
OCR software analyses an image of a page, finds blocks of text, separates lines and characters, and decides, for each shape, which letter it most probably is. Modern engines use machine-learning models and language dictionaries to correct likely errors, and they can output plain text, a searchable PDF or an editable document. It is mature technology, and with clean input it can be very accurate. But it is a statistical guess about pictures, not a copy of the original text, so output should always be proofread.
Where our tools stand today
We want to be direct about this. OCR needs heavy recognition software, and in SmartFileConverters it belongs to the server-based, premium tools: OCR PDF and Image to Text. Those pages describe them, but they are part of the upcoming Pro offering and are not available to run in the free browser tools yet. We do not claim otherwise. What you can do today, free and on your device, is:
- check whether a PDF has real text with PDF to Text;
- create clean scans with Scan to PDF, which turns phone photos of paper documents into a PDF (the pages remain images);
- combine pictures into a PDF with JPG to PDF;
- read and mark up a document with PDF Reader.
If you need OCR right now, use OCR built into your phone's camera or document scanner app, or another OCR program you trust, and be careful about sending sensitive pages to services that upload them.
Why OCR makes mistakes
Even good OCR struggles with predictable things:
- Low resolution. Letters made of a few pixels blur together. Around 300 dots per inch is the usual target for ordinary print.
- Skew and curvature. A tilted page or a curved book spine bends lines of text.
- Poor contrast and shadows. Grey paper, bleed-through from the back of a page and shadows confuse the recognition.
- Unusual fonts and handwriting. Decorative fonts, tiny print and especially cursive handwriting have much higher error rates than clean printed text.
- Layout. Columns, tables, forms, footnotes and text over images can come out in the wrong reading order or merge cells.
- Look-alike characters. The letter O and the digit 0, the letter l and the digit 1, rn and m, are classic confusions that matter a lot in numbers, codes and names.
- Language. OCR must be told which language to expect, and accented or non-Latin scripts need the right language model.
How to get a better scan, step by step
- Light the page evenly from above or from a window, with no shadow of your hand or phone.
- Hold the camera parallel to the page so the text is not trapezoidal, and fill the frame with the page.
- Flatten the paper. Press folded or curled pages under a clear sheet of glass or a book.
- Use enough resolution. A modern phone camera is fine if you are close and in focus. Avoid digital zoom.
- Use Scan to PDF to assemble the pages in order, then check that every page is upright and legible at 100 per cent zoom.
- Do not over-compress. Heavy compression adds blocky artifacts that harm recognition. If you must shrink the file, our guide to compressing a PDF explains the trade-off, and do it after OCR, not before.
Checking the recognised text
Never trust OCR output blindly. Proofread names, dates, amounts, account numbers and addresses against the image. Search for common confusions, such as the digit 0 inside words or the digit 1 where the letter l belongs. For legal, financial or medical documents, treat the recognised text as a working copy and keep the original scan as the authoritative record. If accuracy matters a lot and the document is short, retyping the key figures may be quicker than correcting a mass of errors.
It is also worth deciding whether you need the text at all. If you only need to read the document, keep it as a scan. If you need to find words, a searchable PDF may be enough. If you need to edit, you want a word-processor file, which brings you to the format questions in our guide to PDF, Word and image formats. Choose the least conversion that solves your problem; each step adds the chance of error.
Written by SmarTech Apps. General information only, not legal or professional advice. Questions or corrections? Contact us.