Why are numbers wrong in OCR results?
Digits can resemble letters or each other, especially in condensed fonts, low contrast, glare, blur, and tightly spaced tables.
Digits fail as visual fields, not calculations
Common visual confusions include zero and the letter O, one and lowercase l, five and S, or decimal points and dust. Context may make the OCR output look plausible even when a digit is wrong.
Verify numbers at character level and keep their units, currency, signs, and column headings attached. Capture small numeric fields close up. In a table, rebuild row and column relationships before translating labels.
For a repeatable field-level review, use the OCR confidence check before accepting any amount, date, or unit.
Never ask a language model to repair a total by guessing from surrounding text. For prices, doses, account references, dates, measurements, and safety limits, obtain a clear source and use human review where the consequence is significant.
Audit a number as a complete field
- Compare every digit, decimal separator, minus sign, and thousands separator with the image.
- Keep the currency, unit, date label, row heading, or account-field label attached.
- Use a second close capture when one character controls the value or instruction.
Do not validate numbers by arithmetic alone
A total that adds up can still contain a transcribed value from the wrong row, currency, or date. Arithmetic is a useful secondary check, but the source image remains the evidence. High-impact numeric fields need an independent visual or human confirmation.