Why OCR makes mistakes, and how to get cleaner text

What is OCR actually doing?

OCR (optical character recognition) looks at a picture of a page, finds the lines of text, and decides which character each shape is. It then puts the characters back in reading order. Mistakes come from either step: a shape that could be two different characters, or a page whose reading order is not obvious.

That is why two pages of the same document can come out very differently. One was photographed flat in daylight; the other was taken at an angle under a lamp. The words are the same, the shapes the engine sees are not. To compare tools on pages like yours, use our 15-minute OCR test.

Is the print too small for the image?

This is the most common cause. What matters is not the font size on paper but how many pixels each letter gets in the image. A footnote photographed from across a desk may be only a few pixels tall, and at that size a 6 and an 8, or “rn” and “m”, are the same blob.

Fix: scan at 300 DPI, or move the phone closer so the text fills the frame. Photograph a long page in two halves rather than once from far away.

Is the text faint, ghosted or broken?

Faded thermal receipts, light photocopies, show-through from the other side of thin paper, and printers that drop parts of letters all remove the information OCR needs. When a stroke is missing, the engine fills the gap with its best guess, and a 3 with a broken top can read as a 5.

Fix: capture in bright, even light. For thin paper, put a sheet of black paper behind the page to stop show-through. If the print is truly gone, no tool will bring it back.

Is the photo blurred or taken at an angle?

Motion blur smears every stroke, and a steep angle makes letters at the far edge smaller and squashed. Shadows from your hand or phone make part of the page darker than the rest.

Fix: hold the phone parallel to the page, directly above it, and tap to focus before you shoot. Turn the flash off if it makes a glare spot on glossy paper.

Is the layout confusing the reading order?

Columns, tables, sidebars and captions all ask the same question: which text comes next? An engine can read every word correctly and still interleave two columns, or run a table across its rows. This is a layout error, not a letter error, and it is common on newspapers, forms and bank statements.

Fix: crop to one column or one table before uploading, and keep one page per image. For statements, use a tool built for them; see why statements come out messy.

Is it handwriting or an unusual script?

Neat handwritten print is often read well; joined-up cursive is much harder, because the letters have no gaps between them. Scripts with joined letters or stacked marks, such as Arabic, Urdu and Hindi, have more shapes to tell apart, and decorative or very old typefaces can confuse any engine. Our guide to why some scripts are harder explains each one.

What matters less than people think?

Paper damage on clear print. A coffee stain, a fold, a crease, a watermark or a shadow at the edge usually does not stop OCR from reading clearly printed text. People often re-scan a stained page when the real problem on another page was its tiny footnotes.

What should I try, and in what order?

1. Capture the page again, flat, bright, straight on, with the text filling the frame.
2. Crop to the text you need, cutting away other pages, fingers and background.
3. One page per image, and one column at a time on complex layouts.
4. Run OCR again and compare with the page.
5. Clean up the result. For a single image, AI Polish can fix broken line breaks and obvious misreads. Check names, dates and numbers yourself, because a clean-looking sentence can still have a wrong digit.

When is it time to give up and retype?

When you are fixing more than about one word per line, when the page is handwriting you struggle to read yourself, or when the original is so faded that you are guessing too. OCR saves time when most of the text comes out right; past that point, typing from the page is faster and you will trust the result more.