What Is Optical Character Recognition?

Optical character recognition, commonly abbreviated OCR, is the automatic conversion of images of text into machine-encoded character data. The input is a raster image produced by a scanner, a camera, or a page rendered from a document format, and the output is a sequence of character codes, usually with layout information and per-character confidence scores attached. OCR is the step that makes a scanned page searchable, editable, and machine-readable, and it sits at the front of most document processing pipelines.

Machine reading predates computing. Edmund Fournier d'Albe built the Optophone in 1912 to sound distinct tones for printed characters as an aid to blind readers, Gustav Tauschek patented a template-matching reading machine in 1929, and Emanuel Goldberg demonstrated a Statistical Machine in 1931 that searched microfilm records using photoelectric pattern matching, a patent IBM later acquired. Practical systems arrived with digital computers in the 1950s and were adopted at scale by postal services and banks, where restricted alphabets and purpose-designed typefaces such as OCR-A and E-13B made the problem tractable. Government benchmarking followed: the legacy OCR evaluations and reference data published by NIST supplied the common test corpora on which commercial systems were compared, and the NIST handwritten digit collections were later resampled into the MNIST dataset that became a standard machine learning benchmark.

Preprocessing and Layout Analysis

Recognition accuracy depends heavily on what happens before any character is classified. Binarization separates ink from background, adaptively where illumination is uneven, and noise removal, despeckling, and morphological cleanup follow. Skew estimation corrects the rotation introduced by careless scanning, and dewarping flattens the curvature of a photographed book page. Layout analysis then segments the page into regions and determines reading order, distinguishing columns, headers, tables, captions, and images from body text. Errors at this stage are unrecoverable later: a two-column page treated as one column produces interleaved nonsense regardless of how well individual characters are classified.

Character and Sequence Recognition

Classical OCR segmented text lines into individual characters and classified each one from geometric or statistical features. Segmentation proved fragile for touching or broken glyphs, and modern systems avoid it by treating a whole text line as a sequence. A convolutional network extracts a sequence of feature vectors along the line, a recurrent layer or a transformer models the context, and connectionist temporal classification aligns the network output to a character string without requiring per-character boundaries. A language model or lexicon then constrains the result, which is what allows a system to recover a word whose middle letters are illegible. Handwritten input remains substantially harder than print because of variable slant, ligature, and stroke width, and a review of handwritten optical character recognition surveys the datasets and architectures used for cursive and unconstrained writing.

Scene Text and Constrained-Domain Recognition

Text photographed in the wild introduces perspective distortion, motion blur, uneven lighting, and arbitrary backgrounds, so scene text systems separate detection, which localizes text regions with a bounding box or polygon, from recognition, which reads each cropped region. Automatic license plate recognition is the most heavily engineered instance of this problem, and methods for license plate detection and recognition in unconstrained scenarios rectify oblique views before reading the characters. Contemporary pipelines chain an object detector to a recognition head, an arrangement evaluated in work that integrates YOLO detection with OCR for plate reading. Constrained domains of this kind benefit from a known character set, a fixed field length, and check-digit or format rules that let the system reject implausible readings.

Applications

Optical character recognition has applications in a wide range of fields, including:

• Digitization of archives, historical newspapers, and library collections
• Automated invoice, receipt, and form processing in finance and administration
• Postal sorting through address and postcode reading
• Bank check processing and deposit capture
• License plate recognition for tolling, parking, and traffic enforcement
• Assistive reading technology for blind and low-vision users
• Translation applications that read text through a phone camera