Does MRZ scanning work on damaged or worn documents?

It depends heavily on where and how severe the damage is. The MRZ is designed to be robust.

The OCR-B font's simple, distinct character shapes were chosen partly because they hold up reasonably well even when print quality degrades over years of handling. A passport that's a few years old with some general wear, minor creasing, or slight fading in the MRZ can often still be read successfully, especially with good OCR software that's been trained to handle typical real-world wear rather than assuming pristine, brand-new documents.

More severe damage is a different story. If the physical page is torn through part of the MRZ, if ink has worn away enough that characters are genuinely illegible even to a human eye, or if the document has been water-damaged in a way that's smudged or distorted the text, automated extraction becomes unreliable or impossible for the affected portion.

In these cases, even the best software can't recover data that simply isn't legibly present in the image anymore.

Check-digit validation actually helps here in an indirect but useful way. If a partially damaged MRZ produces a reading that fails its check digit, that's a signal the extraction is unreliable, and a well-designed system will flag the document for manual review rather than confidently presenting a wrong answer as if it were correct.

This matters because a bad outcome in identity verification usually isn't "the scan failed," which is just an inconvenience. It's "the scan silently returned incorrect data that got accepted."

For documents damaged badly enough that MRZ extraction genuinely fails, most systems fall back to visual zone OCR, which reads the printed name and other details directly rather than relying on the MRZ strip, though this loses the additional data-integrity checks that check digits provide.

ScanDoc's approach combines MRZ reading with visual zone OCR precisely so a document with a damaged or unreadable MRZ can often still be processed using the visual zone data, with the system flagging cases where the two sources disagree or where confidence in either extraction is low enough to warrant a human review step.