MMOOZZGG.com

Glossary

What is OCR?

In short

OCR (optical character recognition) is a technology that turns text in a picture, scan or photo into plain text that you can copy and search.

Business and adoption · optical character recognition

MMOOZZGG.com editorsChecked: 1 min read

For a child

When you take a photo of a book page, the phone sees only colored dots, like a mosaic. It does not know there are letters. A special program studies this mosaic, finds letters in it, just as you recognize them in your writing book, and puts them together into words. After that, you can copy the text, find a word in it or have it read aloud.

In plain words

To a computer, a scan of a contract is only a picture: you cannot search it for the word “penalty”. A recognition program finds the lines in the image, splits them into words and characters and guesses which letter is written where. The output is a text file or a PDF in which search works. Printed text on a good scan is recognized almost without errors; handwritten, crumpled or tilted text comes out much worse. A banking app reads a card number from a photo in a similar way.

For business

Recognition is needed wherever paper or scans block automation: a contract archive, incoming invoices, client passports and forms, acts and waybills. Without it, a document assistant will not find an answer in a scanned policy, and a robot will not transfer data from an invoice. Costs depend on the number of pages, the languages, the share of handwriting and whether files may go to an outside service: passport scans fall under 152-FZ, Russia’s personal data law, and are often processed on the company’s own servers. The risks are mixed-up digits in amounts and bank details and a lost table structure. Before launch, run a sample of your own documents and count the error rate in the key fields.

For a specialist

Optical character recognition (OCR) turns an image of text into a machine-readable string. The classic pipeline: binarization and page alignment, finding lines and words (text detection), segmentation, character recognition by a classifier and language post-processing with a dictionary. Ray Smith described the design of the Tesseract engine, which grew out of an HP research prototype, in a paper at ICDAR 2007: line finding, features and classification, and an adaptive classifier. Models on convolutional and recurrent networks with a CTC loss function then spread; they recognize a whole line at once. TrOCR (Li et al., 2021) made recognition fully transformer-based: a pre-trained image transformer encodes the line, a text transformer generates it in word pieces, and the model beat earlier results on printed, handwritten and scene text. Quality is measured by the character error rate (CER) and word error rate (WER). Multimodal models that read a document and extract the fields at once are replacing this stage more and more often.

Example

A legal department scanned an archive of several thousand paper contracts. After recognition, a lawyer finds all contracts with an automatic renewal clause in a minute, instead of leafing through folders.

Related terms

Questions and answers

Can OCR read handwriting? It can, but much worse than print: accuracy depends on the handwriting, the scan quality and the model. For handwritten forms, teams use modern neural network models and keep a human check.
Can you recognize documents without sending them to the cloud? Yes, open engines and recognition models install on your own server. This matters for passports and for medical and HR documents with personal data.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us