What Is OCR? Plain-English Guide to Text Recognition
Learn what OCR (Optical Character Recognition) is, how it works, when it succeeds or fails, and how to turn scans into editable, searchable text for free.
If you have ever photographed a receipt, a page of a book, or a signed contract, and then wanted to copy a line of text out of it, you have wanted OCR. If you have ever scanned a document and been frustrated that it was not searchable, you have wanted OCR. If you have ever received a PDF from a photocopier and could not select the text, you have wanted OCR.
OCR stands for Optical Character Recognition. In plain English: it is the technology that turns pictures of text into actual text.
The problem OCR solves
A photograph of a page is just pixels. To a computer, there is no difference between a picture of the word "hello" and a picture of a sunset. Both are just a grid of numbers representing colour.
A scanner is no different. It captures light, converts it into pixels, and stores them in an image file. The scan of a contract is not a document with text. It is a photograph of a document with text.
Without OCR, that scan is a dead end. You cannot search it, copy from it, translate it, or read it out loud. You can only look at it.
OCR is the step that turns it into something useful. It scans the pixels, recognises patterns that look like characters, and produces a text file that preserves what was in the image.
How OCR works, in six steps
Modern OCR is not magic. It is a specific process that has been refined over 70 years. Here is what actually happens when you feed an image into an OCR engine:
The whole process takes milliseconds per page on a modern device.
- 1
Preprocessing
Grayscale, correct brightness and contrast, remove noise.
- 2
Layout analysis
Find text regions, columns, tables and images.
- 3
Line segmentation
Find individual lines and their baselines.
- 4
Character segmentation
Split each line into individual glyphs.
- 5
Character recognition
Neural network identifies each letter.
- 6
Post-processing
Language model corrects obvious errors.
The whole process takes a fraction of a second per page on a modern device. A 10-page PDF that would have taken a full minute to OCR in 2010 now runs in under 5 seconds on a laptop.
What OCR is good at
- Printed text in any modern font: Times, Helvetica, Arial, Garamond, Roboto, etc.
- Scanned books, invoices, contracts, receipts, letters.
- Screenshots of applications, websites, PDFs.
- Photos taken with a smartphone, as long as the text is in focus and reasonably lit.
- Multi-language documents, including Arabic, Hindi, Chinese and Urdu, with the right language model.
- Structured documents like forms and tables (with a bit of post-processing).
What OCR struggles with
- Handwriting. Neat printing sometimes works; cursive rarely does. Handwriting recognition is a separate problem and requires specialised models.
- Very low resolution. Text below about 150 DPI becomes difficult.
- Heavy skew or rotation. Text that is not level confuses the segmentation step.
- Low contrast. Light gray text on off-white paper, faded ink on old documents.
- Complex mathematical notation, where symbols sit above, below and beside each other.
- Mixed languages in the same passage, where the language model cannot guess the correct word.
When to use OCR
Six common scenarios:
| Scenario | Why OCR helps |
|---|---|
| Digitising old paper documents | Turns scans into searchable, editable text |
| Making PDFs searchable | Adds an invisible text layer behind the scan |
| Extracting text from a screenshot | Copies text from images you cannot select |
| Reading a scanned contract | Find clauses by keyword instead of paging through |
| Translating a foreign document | Get the text into a translator without retyping |
| Automating data entry | Extract numbers from invoices into a spreadsheet |
Three types of OCR
Not all OCR is the same. Three categories are worth knowing:
1. Basic OCR
Extracts raw text from an image. Works well on clean, printed documents. Output is a plain text file with no structure preserved. This is what most free OCR tools do.
2. Layout-aware OCR
Extracts text while preserving the layout: paragraphs, columns, tables. Output can be HTML, Markdown or structured JSON. Better for complex documents but slower and less common in free tools.
3. Handwriting recognition
A much harder problem, solved with dedicated models (Google Cloud Vision, AWS Textract, Microsoft Azure Computer Vision). These work on cursive and messy handwriting but are not free and require an API key.
How to run OCR on a PDF (free, in your browser)
- Open the OCR PDF tool.
- Drop in your scanned PDF. Everything runs locally. Nothing is uploaded.
- Choose the language. The tool supports English, Urdu, Arabic, Hindi, Spanish, French and German.
- Click Run OCR. The first run downloads a language model (one-time, cached).
- Review the extracted text with per-page confidence scores.
- Copy to clipboard or download as .txt.
Free, in your browser. Supports 7 languages.
Run OCR on a PDFWhat confidence scores mean
Every decent OCR engine reports a per-page or per-word confidence score, a percentage showing how sure it is about what it recognised. Here is what to do with it:
- Above 90 percent: clean text. Usually correct as-is. Use with confidence.
- 70 to 90 percent: mostly right. Spot-check proper nouns, numbers and anything you will act on.
- Below 70 percent: expect errors. Re-scan at higher DPI, straighten the page, or use a better language model.
Confidence scores are not perfect. A model can be confidently wrong on unusual fonts or jargon. But they are a useful guide to which pages deserve your attention.
How to get the best OCR result
- Scan at 300 DPI or higher. Anything below 150 DPI is unreliable.
- Straighten the page before scanning. Even 5 degrees of tilt hurts accuracy.
- Use the correct language model. Using English on a Spanish document produces gibberish.
- Improve contrast. Faded print on yellowed paper needs a levels adjustment before OCR.
- Crop out margins and page edges. OCR engines confuse edges with text.
- If a page is one column, do not include another page side-by-side. Column detection is imperfect.
Common myths about OCR
- Myth: "OCR is 100 percent accurate on clean scans." Reality: 95 to 99 percent is realistic on modern fonts. For a 500-word page, that is still 5 to 25 errors.
- Myth: "OCR works on handwriting as well as print." Reality: not even close. Handwriting recognition is a fundamentally harder problem and requires dedicated models.
- Myth: "OCR is expensive or requires a subscription." Reality: modern browser-based OCR is free, fast and private. The model is only 5 to 20 MB per language.
- Myth: "A scanned PDF is the same as a searchable PDF." Reality: a scanned PDF is just an image. A searchable PDF has an invisible text layer added by OCR, which is what makes Ctrl+F work.
The bottom line
OCR is one of the most useful technologies that most people never think about. It is what turned millions of scanned documents into searchable archives, what lets blind users read printed books, what makes digitised legal documents feasible at scale.
For personal use, the free browser tools are now genuinely good. If you have a stack of old scanned documents on your computer, running them through OCR takes seconds per page and makes them searchable forever. That is a permanent improvement to your digital life for a few minutes of work.
Tools mentioned in this article
Frequently asked
What does OCR stand for?
Optical Character Recognition. It is the process of converting images of text into machine-readable text. The word "optical" refers to the fact that the input is a visual image, and "character recognition" is the specific task of identifying each letter, number or symbol in that image.
Is OCR the same as a scanner?
No. A scanner produces an image of a document, a picture. OCR converts that picture into actual text. Some scanners bundle OCR software, and some PDF readers have OCR built in, but they are two separate steps: scan first, then OCR.
Does OCR work on handwriting?
Not well. Standard OCR engines are trained on printed type. Neat handwriting sometimes works, but cursive and messy handwriting are usually beyond them. For handwriting you need a dedicated model like Google Cloud Vision, AWS Textract, or Microsoft Azure Computer Vision. Those are paid services.
How accurate is OCR?
For clean, straight, high-resolution printed text, modern OCR is typically 95 to 99 percent accurate at the character level. Accuracy drops sharply with low resolution, skew, blur, unusual fonts, or handwriting. Every OCR tool reports a confidence score per page so you know which results to trust.
Can OCR handle Arabic, Urdu or Hindi?
Yes, with the correct language model. Tesseract supports most major scripts including Arabic, Devanagari (Hindi), Urdu and Chinese. Using the wrong language model produces gibberish even on clean scans, so always pick the language that matches the document.
Does OCR work on multi-column documents?
Partially. Modern engines have some ability to detect columns and read them in the right order, but the result can still be interleaved if the columns are close together or if the layout is unusual. For important documents, plan on some manual correction after OCR.
Is OCR free?
Yes. Modern browser-based OCR is completely free. The Tesseract OCR engine runs locally in your browser using WebAssembly, so there is no server cost. You get the same accuracy as paid cloud OCR for printed documents. Handwriting recognition requires a paid cloud service.
Is OCR private?
With a browser-based tool, yes. The OCR runs on your device, and your document never leaves your computer. Paid cloud OCR services upload your documents to their servers, which means a copy of the file exists on someone else's infrastructure, a problem for confidential documents.
Does OCR add a text layer to a PDF?
Some tools do. They produce a "searchable PDF" with an invisible text layer behind the scanned image. Others only output plain text. This tool outputs plain .txt, which you can copy or save. If you need a searchable PDF specifically, use a dedicated tool for that step.