Optical Character Recognition (OCR) is the technology that lets a machine "read" like a human: it converts an image of text — a scanned page, a photo of a sign, an invoice — into editable, searchable digital characters. Where the human eye instantly recognizes an "A", a computer initially sees only a grid of light and dark pixels; OCR bridges exactly that gap.
How it works
The classic pipeline chains several stages:
- Pre-processing: binarization (black/white), deskewing, noise removal.
- Segmentation: splitting the image into blocks, lines, words, then characters.
- Recognition: each glyph is matched to templates or classified by a neural network.
- Post-processing: linguistic correction via a dictionary or language model.
Modern systems (TrOCR, Tesseract, end-to-end models) rely on CNNs to extract visual features, paired with sequence architectures (LSTM, Transformers) that decode character sequences directly, without explicit segmentation.
Measuring quality
Accuracy is often measured by the Character Error Rate:
$$ \text{CER} = \frac{S + D + I}{N} $$
where $S$, $D$ and $I$ are substitutions, deletions and insertions, and $N$ is the total number of reference characters. A low CER means a faithful transcription.
| Input type | Difficulty | Example |
|---|---|---|
| Clean printed text | Low | Scanned book |
| Handwriting | High | Notes, archives |
| Natural scene | High | Signs, storefronts |
OCR is now a pillar of digitization: archiving, license-plate reading, instant visual translation, and invoice automation.
From silent pixels to living text: OCR gives machines the ability to read the written world.