What is OCR?
OCR (Optical Character Recognition) is technology that converts images containing text into machine-readable text. A camera or scanner captures an image, and OCR software analyzes the visual patterns to identify letters, words, and sentences.
How OCR works
- Preprocessing: Image is binarized (converted to black and white), deskewed (straightened), and noise-reduced
- Layout analysis: Identifies text regions, columns, and paragraphs
- Character recognition: Each character is compared against trained patterns
- Post-processing: Spell checking and context analysis improve accuracy
OCR accuracy factors
| Factor | Impact on accuracy |
|---|---|
| Image resolution | Higher DPI = better accuracy |
| Font clarity | Printed > handwritten |
| Lighting | Uniform lighting = better |
| Text orientation | Horizontal = best |
| Language | Training data coverage |
| Image noise | Less noise = better |
Best practices for high accuracy
Before taking/scanning the image
- Use 300+ DPI when scanning
- Ensure even lighting without shadows
- Keep text horizontal
- Use high contrast (dark text on white background)
Image types and expected accuracy
- Printed documents: 95–99% accuracy
- Screenshot of digital text: 95–99% accuracy
- Printed signs/labels: 85–95% accuracy
- Handwritten text: 60–80% accuracy (varies widely)
- Artistic/decorative fonts: 50–80% accuracy
Tesseract.js — browser-based OCR
Privatool's Image to Text tool uses Tesseract.js, an open-source OCR engine compiled to WebAssembly. This means:
- OCR runs entirely in your browser
- Images never uploaded to any server
- Works offline after first load (engine is cached)
- Supports 30+ languages
How to extract text from images free
- Go to Image to Text
- Upload image, drag and drop, or paste from clipboard (Ctrl+V)
- Select language
- Click "Extract Text"
- View confidence score and extracted text
- Copy text or download as
.txt
Why OCR fails on some images
Accuracy correlates far more strongly with input quality than with the engine. The main failure modes:
Resolution. OCR needs roughly 300 DPI to segment characters reliably. A phone screenshot of a document is often closer to 72 DPI, and enlarging it afterwards adds no information — the detail was never captured. Photograph the original rather than a screen showing it.
Perspective. A photo taken at an angle produces trapezoidal text. Line-detection assumes roughly horizontal baselines, so skew beyond a few degrees degrades results sharply. Shoot straight down, or de-skew before running OCR.
Contrast and shadows. Binarisation splits pixels into ink and background using a threshold. A shadow falling across the page shifts local brightness so that one region's ink is darker than another's paper, and a single global threshold cannot separate them. Diffuse, even lighting matters more than brightness.
Compression artefacts. Heavy JPEG compression produces ringing around high-contrast edges — precisely where letterforms live. Prefer PNG for screenshots of text.
Fonts. Clean serif and sans-serif body text is what engines are trained on. Handwriting, decorative display faces, and stylised logos perform poorly to the point of being unusable.
Where the errors cluster
OCR mistakes are systematic rather than random, which makes proofreading faster once you know the pattern:
| Confused | With | Context |
|---|---|---|
0 |
O |
Serial numbers, codes |
1 |
l, I |
Sans-serif fonts especially |
5 |
S |
Low resolution |
rn |
m |
Tight kerning |
cl |
d |
Narrow spacing |
Numbers and alphanumeric codes are the highest-risk output, because there is no dictionary to fall back on and no context to disambiguate. Always verify OCR'd account numbers, invoice references, and IDs by eye.
Layout is a separate problem
Recognising characters is only half the task. Reconstructing reading order is the other half, and it is where multi-column documents fail. An engine that reads across the page rather than down each column produces text that is individually correct and collectively meaningless.
Tables are harder still: cell boundaries carry meaning that plain text cannot express. For a table, expect to rebuild the structure by hand.
Practical workaround: crop to a single column and run OCR on each separately.
Why browser OCR is worth the trade-off
Tesseract compiled to WebAssembly is slower than a cloud OCR service and somewhat less accurate on difficult inputs. In exchange, the image never leaves your device.
That trade matters because of what people actually run OCR on: passports and ID cards, invoices, medical letters, contracts, bank statements. This is among the most sensitive categories of document, and the standard workflow is to upload it to a third party. For a photographed ID document, local processing is worth accepting some accuracy loss.
The first run downloads a language model of several megabytes, then caches it. Subsequent runs start immediately and work offline.
Frequently asked questions
Can OCR read handwriting?
Standard OCR, essentially no. Handwriting recognition (ICR) is a different technology requiring different models, and even the best cloud services struggle with unconstrained cursive.
Which language should I select?
The one matching the document. Selecting the wrong language forces the engine to fit text to the wrong character set and dictionary, which degrades output badly. Multi-language documents need a model trained on those languages together.
Why is my output full of random symbols?
Usually the image is too low-resolution, or the language model does not match the script. Try a higher-resolution capture first — it fixes this more often than any setting.
Can OCR preserve formatting?
Plain-text OCR does not. Bold, italics, and layout are lost. Some engines emit hOCR or ALTO with positional data, but reconstructing a formatted document from it is a separate job.
Is OCR accurate enough for legal or financial documents?
Not without human verification. Even at 99% character accuracy, a page of 2,000 characters contains around 20 errors — and in a financial document a single wrong digit matters. Treat OCR as a fast first draft, never as authoritative.