Why OCR gets some words wrong, and how to get a cleaner result

How OCR reads a page

OCR does not read the way you do. It splits the page into blocks, lines and words, compares the shape of each word with what words look like in the chosen language, and picks the most likely reading. It never knows it is right; it chooses the best match it can find.

That is why the quality of the picture matters so much. At around 300 dots per inch, ordinary body text has enough detail for each letter to be unmistakable. At half that, an “e” and a “c”, or “rn” and “m”, start to look alike, and the guesses get worse.

The usual causes of mistakes

• Blur. A photo taken while moving, or focused on the wrong part of the page. Edges smear and letters run together.
• Shadows and glare. The shadow of a hand or phone, or a shiny page reflecting a lamp, wipes out part of a line.
• Small or faint type. Footnotes, faded receipts and light grey text have too few dark dots to work with.
• A curved or crooked page. The middle of a thick book bends the lines; a page photographed at an angle squeezes one side.
• The wrong language. A French page read as English produces confident nonsense, because the shapes are matched against the wrong words.
• Handwriting. OCR is built for printed letters. Neat capitals sometimes work; joined-up writing usually does not.

Getting a better picture

1. Lay the page flat and photograph it straight on, with the text filling most of the frame.
2. Use even daylight, or turn the lamp so no shadow falls across the page. Switch the flash off for shiny paper.
3. Tap the screen to focus on the text, hold still, and check that the photo is sharp before you upload it.
4. Take a screenshot at the size you read it; do not zoom out first.
5. On a scanner, choose 300 DPI. Colour is not needed for text: greyscale is just as accurate and much smaller.

Checking the result

Read the result against the picture before you rely on it, and pay most attention to what OCR cannot sense-check: names, numbers, dates, amounts and codes. A misread letter inside a word is usually obvious; a misread digit in an account number is not. iBuildPDF says which language the text was read in, and if that is wrong, choose the right one and run it again. There is more on this in how to check a converted document.