A QR alignment pattern searched as 1:1:1 runs along image rows is missed on rotated codes

Fact

ZXing-style detectors look for the alignment pattern by scanning image rows for light-dark-light runs, each within half a module of the estimated module size. A row crosses a rotated pattern's rings diagonally, which stretches every run by up to √2 at 45°. Add a little blur and the real pattern fails the test, while a chance 1:1:1 run in the data area passes and is used for the perspective transform.

Matching the whole 5×5 template (dark ring, light ring, dark centre), sampled at steps of the symbol's own module vectors, does not depend on rotation or skew.

Why it matters

The alignment pattern is the fourth point of the perspective fit. Losing it, or locking onto a look-alike a few modules away, makes every tilted code undecodable. The finder patterns are all found, so it looks like a sampling or error-correction bug.

How to apply

• Take the module vectors from the three finder centres (or from the current homography), then slide the template over a window around the predicted centre. Score by mismatches, and take the centroid of the lowest plateau.
• Keep a few candidates and confirm each by a full decode, because Reed-Solomon is the only reliable judge of a fit.
• On large versions, find the alignment patterns nearest the finders first and refit a least-squares homography after each, so every prediction is a short extrapolation.