Pick up a cereal box, a train ticket, or a restaurant menu and there is a good chance a QR code is looking back. Look closely and the same three shapes appear on almost every one: a thick black square, a white ring inside it, a smaller black square at the centre, stacked in the top-left, top-right, and bottom-left corners. They look like design flourishes. They are not. They are the reason a phone can read the code held sideways, upside down, reflected in a mirror, or photographed from across a café at an awkward tilt.
The squares are called finder patterns, and they are specified in ISO/IEC 18004, the international standard that defines the QR code symbology. They were designed by a team at the Japanese automotive parts manufacturer Denso Wave, led by engineer Masahiro Hara, and published in 1994 to track components on Toyota production lines — a history Japan’s own patent office has documented in its account of the invention, which credits Hara’s team directly.
The barcodes used at the time could hold about 20 characters and needed to be aligned reasonably straight with the scanner. Hara’s team wanted a code that held hundreds of times more data and could be read fast, from any angle, by a factory scanner sweeping across a moving part.
Orientation was the hard part. And the answer was those three squares.
What the finder patterns actually do
Each finder pattern is a nested set of three concentric squares in a fixed ratio: a seven-module black square, a five-module white square set inside it, and a three-module black square — solid, not hollow — at the centre. A “module” is the smallest black or white cell in the code. Scanned along any straight line that passes through the middle of one of these patterns — horizontally, vertically, or diagonally — the sequence of black and white runs comes out in the ratio 1 : 1 : 3 : 1 : 1.
Each finder pattern is also wrapped in a further one-module-wide light border, called a separator, that keeps its black edge from touching the black data modules next to it. Without that thin gap, a scanner sweeping across the boundary could not tell where the finder pattern’s fixed shape ends and the unpredictable data region begins.
That ratio is the point. It is rare in ordinary imagery, and it survives rotation. A scanner does not need to know which way up the code is. It sweeps lines across the image looking for that specific run-length pattern, and wherever it finds the ratio in both a horizontal and a vertical sweep at the same location, it has almost certainly found a finder pattern. The QR standard describes exactly this detection method in its reference decoding algorithm.
Three finder patterns, one missing corner. That asymmetry is deliberate. Once the scanner has located the three squares, it knows which corner is which: the bottom-right corner, the one without a finder pattern, tells the decoder how the code is rotated relative to the camera. Upside down, the three squares still form an L; the empty corner is just in a different place. The scanner rotates its internal model of the code to match, and reads the data in the correct order.
Why it takes less than a millisecond
The finder pattern search is fast because it is a one-dimensional problem disguised as a two-dimensional one. Instead of analysing every pixel and trying to understand the whole image, the decoder walks across rows and columns of the captured frame looking for that 1 : 1 : 3 : 1 : 1 run-length signature.
Modern phones do this on a downsampled greyscale version of the camera frame, often only a few hundred pixels wide, and the run-length check is a handful of integer comparisons per candidate.
Locating the three corners answers where and which-way-up. A separate feature answers how big: a single row and column of alternating black and white modules, called the timing pattern, run between the finder patterns, and counting those alternations tells the decoder exactly how many pixels make up one module — the unit it needs to lay an accurate grid over everything else in the symbol.
On a current smartphone processor, locating the three finders and computing the perspective transform typically takes well under a millisecond of the total decode time. The heavier work — reading the data modules, applying Reed–Solomon error correction, decoding the payload — happens after the orientation is already fixed.
This is why the experience feels instant. By the time the viewfinder has drawn a frame around the code, the phone has already worked out where the corners are, how tilted the code is, and where every data cell should fall.
Handling tilt, curvature, and the 45-degree photo
The three squares also solve a subtler problem: perspective. A QR code printed on a curved coffee cup or photographed at 45 degrees is no longer a square in the image. It is a quadrilateral, sometimes a warped one. Three known points, plus the geometry of a fourth inferred corner, are enough for the decoder to compute a homography — a mathematical mapping that undoes the perspective distortion and turns the skewed image back into a regular grid.
For larger codes, the standard adds smaller alignment patterns (single 5×5 nested squares) at fixed positions inside the grid, a refinement Denso Wave introduced with the symbol format it calls Model 2. These give the decoder extra anchor points so that codes bent around a bottle, or printed on a wrinkled label, can still be reconstructed cell by cell. Version 1 QR codes, at 21×21 modules, do not need any. From version 2 upwards, alignment patterns start appearing, and the largest version — version 40, at 177×177 modules — uses 46 of them.
None of this works without a plain margin around the whole symbol. The QR Code specification calls for a blank quiet zone at least four modules wide on every side, so the finder-pattern search has clean space to work in and doesn’t mistake nearby text or graphics for the edge of the code.
The error correction layer picks up whatever the geometry cannot. QR codes carry Reed–Solomon codewords with four selectable levels of redundancy, from about 7 per cent (level L) up to about 30 per cent (level H). A code at level H can lose almost a third of its area — including, famously, the space behind a logo pasted in the middle — and still decode. That is why brand-stamped QR codes work at all: the designer punches out a chunk of data cells and lets the error correction fill in the missing information.
What the record actually shows about the design
The often-repeated story is that Hara chose the 1 : 1 : 3 : 1 : 1 ratio after his team analysed printed materials — newspapers, magazines, business documents — to find a run-length pattern that almost never occurs naturally, so the finder would not be confused by background imagery.
Denso Wave’s own history of the QR Code credits this reasoning directly to Hara, describing an exhaustive survey of the ratio of black to white areas across fliers, magazines, and cardboard boxes, undertaken to find the shape least likely to turn up on ordinary printed matter. Hara has described the same reasoning in interviews with outlets such as SBS. The claim that the ratio is the rarest sequence in printed matter comes from Hara directly; it has not been independently reproduced as a study, and should be read as the designer’s account rather than a peer-reviewed finding.
What the standard document itself makes explicit is narrower: the ratio is defined, the detection algorithm is defined, and the ratio is chosen so that a scan line at any rotation through the centre of the pattern produces the same signature. That property is what makes the code rotation-invariant. The rest — how rare the ratio is in the wild — is the design story around it.
The consequence you never noticed
Every time a phone reads a QR code, the sequence is the same. The camera grabs a frame. A downsampled copy is scanned for the 1 : 1 : 3 : 1 : 1 signature in rows and columns. Three matches form an L. The missing corner tells the decoder which way is up. A perspective transform snaps the grid back into a square. Error correction cleans up the rest. Only then does the payload — the URL, the boarding pass reference, the Wi-Fi credentials — get read.
The three corner squares carry no user data at all. They exist so that the other 400-odd, or 30,000-odd, cells that do carry data can be found in any orientation, on any surface, in under a millisecond. It is the same category of quiet infrastructure as the DNS root servers or the file-system pointers that survive a “delete”: a piece of the technology that carries the whole thing but is almost never noticed.
Denso Wave still holds the QR code patents but has chosen not to enforce them, which is why the format spread from Toyota’s assembly lines to menus, tickets, and payment terminals worldwide. The design has not changed meaningfully in 30 years. It did not need to.