optical character recognition

Optical character recognition turns an image of text — a scanned page, a photographed sign, a PDF of a fax — into machine-readable characters you can search, copy, and edit. The goal is to recover the symbolic content (the letters and digits) from the picture of it. OCR is one of the oldest applied recognition problems (commercial systems date to the 1950s) and a clean illustration of the classical recognition pipeline, because for decades it was built as an explicit chain of hand-designed stages rather than a single learned model.

The classical pipeline runs roughly as follows. First, preprocessing cleans and normalizes the image: binarization (separating ink from paper, often with an adaptive threshold), deskewing (rotating a tilted scan upright), and noise removal. Then layout analysis and segmentation carve the page into columns, text lines, words, and finally individual characters — the segmentation step, deciding exactly where one character ends and the next begins, is historically the hardest and most error-prone part, especially for touching or cursive letters. Next, feature extraction turns each character image into a descriptor robust to font and size — early systems used template matching or stroke and zoning features, later ones used gradient features. Finally a classifier (kNN, SVM, neural net) maps each descriptor to a character label. A crucial last stage is language modeling: a dictionary, character n-gram statistics, or grammar corrects errors using context, so that rn is not misread for m and l is not misread for 1 when the surrounding word makes the right answer clear.

OCR matters both as a hugely useful technology (digitizing books and archives, reading checks and forms, license plates, accessibility for the blind, document automation) and as a microcosm of how recognition evolved. Modern OCR collapsed much of the classical pipeline into learned models: convolutional and recurrent networks trained with connectionist temporal classification (CTC) recognize a whole text line at once without explicit character segmentation (the segmentation-free approach), and current systems use attention-based sequence models and transformers (for example Microsoft's TrOCR), reaching near-human accuracy on clean printed text and handling photographed scene text, handwriting, and many scripts. The classical stages survive conceptually inside these systems and remain instructive for understanding where errors come from.