Optical character recognition (OCR)

Optical character recognition turns images of text into actual text. Given a photo, scan, or video frame, an OCR system first finds where text appears and then recognizes the characters and words, outputting a machine-readable transcription. It handles everything from clean printed documents to messy text in natural scenes, like signs and labels.

It is what makes the words inside images searchable, editable, and usable by other software.

Modern OCR generally works in two stages: text detection locates regions containing text, often with the same techniques as object detection, and text recognition transcribes the characters within each region, typically with sequence models. Scene text adds difficulty from varied fonts, angles, lighting, and backgrounds, so models are trained on diverse, well-annotated text data.

Huge amounts of information exist only as text inside images and documents, and OCR is what unlocks it for search, automation, and analysis. In visual AI it connects image understanding to language, and its quality depends directly on the OCR annotation and text annotation used to train it.

Reading text that appears in natural images, such as signs and labels, which is harder than document OCR because of varied fonts, angles, and backgrounds.

Detecting where text is located in the image, then recognizing and transcribing the characters within those regions.

OCR annotation, Text annotation, Object detection, Image classification.