Optical character recognition (OCR) is a technology that extracts text from images, scanned documents, or photographs of text, converting pixels into machine-readable and machine-editable characters. It sits at the intersection of Computer vision and Natural language processing, since a working system must both locate and segment glyphs in an image and interpret them as symbols in a language. OCR was one of the first commercially useful applications of pattern recognition and predates the modern Deep learning era by decades.
History
Early OCR systems date to the 1950s and 1960s, when devices such as David Shepard's "Gismo" and Ray Kurzweil's reading machines for the blind (1976) used template matching and hand-engineered features to recognize printed fonts. Through the 1980s and 1990s, OCR software such as Tesseract, originally developed at Hewlett-Packard and later open-sourced by Google, relied on statistical classifiers and feature extraction pipelines built largely with Symbolic AI and classical machine learning rather than the layered representation learning that later became dominant. Accuracy on clean, typed text reached commercial usability by the 1990s, but handwriting and cluttered scenes remained difficult problems for another two decades.
Deep learning era
The introduction of convolutional neural networks and later recurrent and sequence-to-sequence architectures dramatically improved OCR accuracy, especially for handwriting and text embedded in natural scenes, such as street signs or product labels. Architectures combining a CNN for visual feature extraction with an LSTM or attention-based decoder became standard through the mid-2010s. As with much of Computer vision, progress tracked the broader shift from hand-engineered pipelines toward end-to-end neural networks trained on large labeled datasets.
The vision-language model era
Since the early 2020s, general-purpose vision-language models and multimodal systems built on the Transformer (architecture) architecture have absorbed much of OCR's functionality. Models with vision capability, including Gemini and GPT-4 variants, can read text embedded in an image as part of a broader visual understanding task, without a dedicated OCR pipeline, and often outperform specialized OCR systems on documents that mix text with diagrams, tables, or handwriting in context. This has blurred the boundary between OCR as a discrete task and general document understanding, though purpose-built OCR engines remain in wide use for high-throughput, low-latency, or offline document processing, where a large multimodal model would be costlier or slower.
Applications and evaluation
OCR underlies document digitization, searchable PDF creation, license-plate recognition, receipt and invoice processing, postal mail sorting, accessibility tools for blind and low-vision users, and the ingestion pipelines that feed text into search engines and large language model training corpora. Systems are typically evaluated on character error rate and word error rate against benchmark datasets of scanned or photographed text; performance varies widely by language, script, print quality, and handwriting style, with cursive handwriting and low-resource scripts remaining comparatively weak points. As a component task, OCR quality is usually reported on dedicated document and scene-text datasets rather than general-purpose benchmarks.
Limitations
OCR errors can propagate into downstream systems, such as search indexes or automated data entry, and specialized attacks exist that exploit visually similar characters to evade OCR-based filters. Multimodal models that perform OCR implicitly can also hallucinate text that resembles but does not match the source image, a failure mode distinct from classical OCR's segmentation errors.