Optical character recognition (OCR)
Software that turns an image of text, such as a photo of a receipt, into characters a computer can search and process.
In short
- OCR reads characters. It does not know which number is the total or which word is the merchant.
- Faded thermal paper, creases, glare and odd fonts are the usual causes of errors.
- Data sent directly from the point of sale skips OCR entirely, and with it the reading errors.
Optical character recognition, or OCR, converts pictures of text into text. Point a phone at a receipt and OCR produces the words and numbers printed on it, in roughly the order they appear.
Reading is not understanding
On its own, OCR returns a block of text. It does not know that "43.70" is the total, "5.12" is tax and "4501" is a store number. A second step, often a machine-learning model, has to interpret the text and assign each value to a field. When people say a receipt app "uses OCR", they usually mean both steps together.
Where it goes wrong
- Image quality: blur, shadows and glare hide characters.
- Paper: thermal receipts fade. Older ones may be partly blank.
- Look-alike characters: 5 and S, 0 and O, 1 and l are frequent swaps, and a single wrong digit changes an amount.
- Layout: columns and wrapped lines can be read in the wrong order.
OCR works well enough that a person only needs to check exceptions, but it is never perfect. The only way to remove reading errors completely is to receive the receipt as data from the system that issued it.
Worked example
One digit, two outcomes
A crumpled parking receipt reads $18.50. The OCR returns "$1B.50", then the interpretation step guesses $13.50. A total check against the $18.50 card charge flags the mismatch, and the employee corrects it in a few seconds. Without that check, the books would be $5 off and nobody would know why.