Blurred writing, and whether four characters can be read

On the front page a result is judged by looking at it. Here it is not. Nobody arriving with a photographed enrolment form, a certificate or a delivery label wants a handsome picture; they want four characters, and four characters either resolve or they do not. That changes what the pass is for and it changes what an honest page is obliged to say about the output.

Flat printed matter is, on paper, the best case this method has. A page is a plane, so every part of the frame carries the same smear and one measurement describes all of it. Strokes are high-contrast against paper, which means there is real energy to gather back up rather than a soft gradient. And a hand-held exposure indoors — the shutter open long enough for a few millimetres of movement — is the exact failure this arithmetic is shaped for.

It is also a bet. The learned pass this domain is built around was trained on natural scenes, not on documents, and whether what it learned about edges transfers to a dense field of printed strokes is untested here. That is written down rather than assumed, and it is the first thing that should be measured when an export is published.

• A reconstructed glyph can be confidently wrong. The pass returns the sharpest strokes consistent with what was recorded, and several characters are consistent with the same smear. Nothing on screen marks which ones were guesses.
• A number read off a result is not evidence. Check it against the document, the original photograph or whoever issued it before it is used for anything that matters.
• A tilted page carries more than one smear and is measured as though it carried one. Crop first.
• A redacted region is a block grid, and blocks are not blur. The rail prints the period it found; that number is the reason not to keep pressing.
• Low resolution is not blur either, and re-photographing the page will beat every pass on this domain in a fraction of the time.

None of that stops the pass being worth running on a badly shot form. It sets what the output is: a legible candidate, produced from what the sensor actually caught, which a person still has to confirm.

Why is a wrong character worse than a blurry one?

Because a blurry character announces itself and a wrong one does not. Given a smeared shape, the pass produces the sharpest arrangement of strokes consistent with what the sensor recorded, and a 3 whose lower bowl has been spread across four pixels is consistent with an 8. The result is clean, legible and occasionally false, which is precisely the failure that costs somebody an account number. Anything read off a pass here has to be checked against something else before it is used.

Will this read the text for me?

No. Turning the characters on a page into a string is optical character recognition, it is a separate operation with separate failure modes, and it belongs to pictureeditor.com along with any chain of two operations. This page makes strokes legible to a person. What that person then does with them is not something this domain touches.

My scan is low resolution rather than blurred. Does a pass help?

No, and the measurement will usually say so. Low resolution means the strokes were never sampled finely enough to exist in the file; blur means they were sampled and then spread. Only the second has anything to gather back together. Re-photographing the page at a shorter distance adds real samples, and it takes less time than any of this.

What should I do differently when re-shooting a document?

Put the page flat, brace the phone against something solid rather than holding it out, and take the exposure away from the shadow your own body casts over the page. Most indoor document photographs fail because the shutter stayed open long enough for a hand to move a few millimetres, and a surface to lean on removes the cause instead of arguing with the result afterwards.

Does the pass handle a page that is only blurred in one corner?

Poorly. The measurement reads one smear for the whole frame, which is right for a page photographed flat and wrong for one that was tilted, where the far edge moved further across the sensor than the near edge did. Cropping to the part you actually need before opening it gives the measurement a region with one consistent failure in it, and that is the single most useful thing you can do here.

Is coloured or reversed text different?

Only in how much contrast survives. The pass works on each of the three channels separately, so white-on-dark behaves the same as dark-on-white. What hurts legibility is low contrast to begin with — pale grey on white, or a watermark printed behind the line you want — because the smear has less to spread and there is correspondingly less to pull back.

The general case is on the front page and reads a photograph rather than a page. The weights page carries the licence chain and the provenance. The long read on what can be reconstructed is the one to start with if the question underneath yours is whether any of this works at all.

• The passOne photograph in, one pass over it, one photograph back.
• The weightsWhich network, whose licence, which corpus, and what it costs to fetch.
• What can be reconstructedLearned priors, the GoPro corpus, and reading a result sceptically.