Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Optical Character Recognition

nounComputer Visionalso OCR, text recognition, handwriting recognition

In one line

Optical character recognition (OCR) is the technology that converts images of printed, typed or handwritten text into machine-readable text.

Optical character recognition (OCR) is the technology that converts images of text into machine-readable text. A scanned page, a photograph of a sign, a PDF made of pictures or a handwritten note is only a grid of pixels to a computer. OCR finds the characters in that grid and outputs actual text that can be searched, copied, edited, translated and stored. It is one of the oldest practical applications of pattern recognition and of image recognition, and it is now a routine part of document handling, accessibility tools and mobile apps.

Overview

OCR sits between the physical world of paper and the digital world of data. Companies use it to digitize archives and process invoices, libraries use it to make books searchable, and phones use it to read text in the camera view. The task has several variants.

  • Printed text OCR. Recognizing typed or printed text in documents. This is the most mature form and can reach very high accuracy on clean scans.
  • Handwriting recognition. Reading handwritten text. It is harder because styles vary greatly. It is called offline when working from an image and online when it uses pen strokes captured as they are written.
  • Scene text recognition. Reading text in natural photographs, such as street signs, shop fronts and product labels, with varied fonts, angles and lighting.
  • Intelligent document processing. Going beyond reading characters to understanding layout and extracting fields such as dates and totals.

History

The idea of machines that read dates back to the early 20th century, with devices to help blind readers and to read telegraph text. In the 1970s, Ray Kurzweil developed omni-font OCR, which could recognize text in many typefaces, and in 1976 introduced the Kurzweil Reading Machine to read books aloud for blind people [1]. Banks and postal services adopted automatic reading of checks and addresses. Open-source engines such as Tesseract, begun at Hewlett-Packard and released as open source in 2005, made OCR widely available [2]. Early systems matched characters against templates or extracted hand-designed features. Since the 2010s, neural networks, including convolutional and recurrent models, and more recently transformers, have raised accuracy, especially on difficult images and handwriting. A landmark early neural success was LeNet reading digits on checks.

How OCR works

A traditional pipeline has several stages, and modern systems merge some of them.

  1. Image acquisition. A scanner or camera captures the page.
  2. Preprocessing. The image is cleaned: converted to black and white or grayscale, deskewed so that lines are horizontal, denoised and corrected for uneven lighting.
  3. Layout analysis. The system finds columns, paragraphs, tables, images and reading order. This step is a form of image segmentation.
  4. Text detection and line segmentation. Regions of text are located, in the manner of object detection, and split into lines and words.
  5. Recognition. A model converts each line or word image into characters. Older systems classified individual characters, while newer ones read whole lines using sequence models, so that no explicit character segmentation is needed.
  6. Post-processing. Dictionaries and language models correct likely errors, for example fixing a misread “rn” that should be “m.” Output may preserve formatting in a searchable PDF or structured file.

End-to-end deep learning models can read text directly from an image and output a transcription, and recent vision-language models can read and interpret documents together; see multimodal AI.

Evaluation

Accuracy is usually measured by character error rate and word error rate, which count the insertions, deletions and substitutions needed to turn the OCR output into the correct text, relative to its length. Results depend heavily on input quality. Clean modern print can reach well above 99 percent character accuracy, while degraded historical documents, low-resolution photos and cursive handwriting can be far lower. Because errors may fall on important items such as names and numbers, applications often check key fields separately.

Applications

  • Document digitization. Converting paper archives, books and newspapers into searchable text, including cultural heritage collections.
  • Business automation. Extracting data from invoices, receipts, forms and contracts, often with validation by a person.
  • Banking and mail. Reading checks, postal addresses and identity documents.
  • Accessibility. Reading text aloud for people with visual impairments, together with speech synthesis.
  • Translation. Reading signs and menus through a phone camera and passing the text to machine translation.
  • Transportation. Reading vehicle license plates for tolling and parking, which raises privacy questions.
  • Search and data mining. Making scanned text available to text mining and semantic search.
  • Education and research. Transcribing manuscripts and historical records.

Challenges

  • Image quality. Blur, noise, low resolution, shadows and stains lower accuracy.
  • Fonts and handwriting. Unusual typefaces, cursive writing and mixed styles are difficult.
  • Complex layouts. Tables, multi-column pages, footnotes and mixed images require careful layout analysis.
  • Languages and scripts. Scripts such as Arabic, with connected letters, or Chinese, with thousands of characters, need specialized models, and low-resource languages lack data.
  • Historical documents. Old print, faded ink, spelling variants and damage reduce accuracy.
  • Privacy and security. Digitized documents may contain personal data that must be protected.
  • Hallucinated text. Generative models that read documents can sometimes output plausible text that is not on the page, so critical uses need verification.

Document understanding beyond OCR

Reading characters is only the first step in handling real documents. Document understanding systems use layout, font size, position and visual cues to work out which text is a title, a table cell, a form label or a total. Models that combine text, layout and image information can extract named fields from invoices, tax forms and identity documents, often with no template for each supplier. Output is normally checked against rules, such as whether line items add up to the total, and uncertain fields are passed to a person for review. Related tasks include table structure recognition, key-value extraction and answering questions about a page; see visual question answering.

Improving OCR results

Results improve markedly when the input is good. Scanning at a resolution of 300 dots per inch or higher, keeping pages flat and well lit, and avoiding compression artifacts all help. Choosing the right language setting, supplying a domain dictionary for specialist terms and adding a spelling correction step reduce errors. For historical material, models fine-tuned on similar print or handwriting give much better results than general engines, and crowdsourced correction by volunteers is a common way to reach archive-quality text.

Accuracy, access and ethics

OCR has done much to widen access to information, from reading aloud for blind users to making old newspapers searchable. At the same time, errors can propagate when uncorrected text is stored as if it were reliable, and automated reading of license plates, documents and handwriting can enable surveillance if poorly governed. Good practice is to keep the original image alongside the text, to record the confidence of each recognition, to limit the retention of personal data and to tell users when a text is machine-read and may contain mistakes.

FAQs

What is OCR? Optical character recognition is technology that converts images of printed or handwritten text into machine-readable text.

How accurate is OCR? On clean printed text it can be very accurate, but accuracy drops for poor scans, unusual fonts, handwriting and complex layouts.

What is the difference between OCR and handwriting recognition? OCR usually refers to printed text. Handwriting recognition deals with the greater variability of handwriting, and it is generally less accurate.

What is scene text recognition? It is the reading of text in natural photographs, such as signs and labels, where fonts, angles and lighting vary widely.

Is OCR free to use? Open-source engines such as Tesseract are free, and many commercial services and phone apps offer OCR as well.

Key facts

  • Ray Kurzweil introduced the Kurzweil Reading Machine in 1976, a device that used omni-font OCR to read printed text aloud for blind users1.
  • Tesseract, an open-source OCR engine originally developed at Hewlett-Packard in the 1980s and 1990s, was released as open source in 2005 and later sponsored by Google2.
  • The IAM Handwriting Database and the MNIST digit set are among the widely used data sets for evaluating handwriting recognition3.

Related concepts

Quick checkWhat is optical character recognition?Show answer

Optical character recognition (OCR) is technology that converts images of text, such as scanned pages, photographs of signs or handwritten notes, into machine-readable text that can be searched, edited and stored.

Sources

  1. National Federation of the Blind, The Kurzweil Reading Machine, https://nfb.org/about-us/history
  2. Ray Smith, An Overview of the Tesseract OCR Engine, Ninth International Conference on Document Analysis and Recognition, https://research.google/pubs/an-overview-of-the-tesseract-ocr-engine/
  3. Urs-Viktor Marti and Horst Bunke, The IAM-database: an English sentence database for offline handwriting recognition, International Journal on Document Analysis and Recognition, https://doi.org/10.1007/s100320200071

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue