Date stamps on scanned prints

Between about 1985 and the end of consumer film, a great many compact cameras had a quartz date back: a small module behind the film gate holding a clock and an LED. When the shutter fired, the module exposed the date directly onto the emulsion in the bottom corner of the frame. It was a selling point. It was also, with hindsight, a permanent alteration to every photograph anybody took for fifteen years.

Two consequences follow from the fact that the date was exposed onto the negative rather than printed onto the paper, and they explain most of what is difficult about taking one off. The first is colour: the LED was red or amber, colour negative film inverts, and what comes back is the dense orange-yellow every one of these imprints shares. It is a very unusual colour for a photograph to contain otherwise, which is the whole reason an automatic scan is possible at all. The second is that the digits are not a layer. Light hit the emulsion, silver formed, dye clouds developed around it; the imprint has grain, it has a soft halo where light scattered sideways, and at the edges it is genuinely mixed with the scene behind it. There is nothing to peel off.

Painting over seven digits by hand at a comfortable zoom takes maybe twenty seconds and produces a region with soft, uneven edges that either misses parts of a stroke or claims a good deal of picture that did not need filling. Neither is fatal, but both are avoidable, because an imprint has a property almost nothing else in the frame has: it is a colour that is not in the scene.

Selecting by similar colour works by flooding outward from the pixel you point at, taking every neighbour within a tolerance. The measure matters. In plain red-green-blue terms, a fifteen-step difference in a dark blue is invisible and the same fifteen steps in a light yellow is a different colour entirely, so one tolerance number would mean two different things in two parts of the same picture. Converting to a perceptual space first — the one designed so that equal numerical distances look like equal colour differences — makes the tolerance slider mean roughly the same thing everywhere. Under about five is “the same colour”; fifteen is “the same material”; forty is “the same half of the photograph”.

There is a second test running alongside it, and it is what stops the selection escaping. A candidate pixel has to be close to the colour you clicked and close to the pixel it was reached from. Without the second test, a sky that shifts imperceptibly from one side of the frame to the other is one continuous chain of tiny steps, and a single click swallows the whole thing.

The most common complaint about an imprint removal is not that the digits are still there. It is that the digits are gone and a faint orange bloom is where they were. That is the halo: the region covered the strokes but not the light that scattered around them, so the fill averaged intact picture together with imprint-coloured pixels it had been told were intact.

Growing the region by two pixels before running fixes it nearly every time, and it costs nothing, because the pixels being sacrificed were contaminated anyway. It is worth doing by default rather than as a remedy. Feathering is the other half: a hard boundary between filled and original pixels is visible at any zoom even when both sides are individually convincing, because what the eye reads is the discontinuity and not the content. Two pixels of soft edge is enough on most scans.

• Over sky, sand, water, grass, snow or a plain wall, an imprint removal is reliably invisible. There is real picture within a few pixels in every direction and the strokes are narrow.
• Over fine repeating detail — brickwork, a patterned carpet, foliage at distance — switch to the patch search. Averaging flattens texture; copying does not.
• Over a strong edge that crosses the digits, expect the edge to come out slightly wrong, and consider whether a crop is the better answer.
• Over a face or over lettering, do not. Both come back as plausible nonsense, and on a photograph of a person that is worse than the date.

When a picture opens on the date-stamp page, the bottom quarter of the frame is examined for pixels in a narrow orange-to-yellow hue window, saturated past about a third and bright enough to be a light source rather than a shadow. Those are grouped into connected blobs. A blob is kept if its height is between about one and seven per cent of the frame height, if it is taller than it is wide or close to square, and if it fills between a fifth and four fifths of its own bounding box — because a digit made of straight segments fills part of its box, a blob of sunlight fills all of it, and a speck of noise fills almost none. What survives is then grouped into runs that share a baseline and sit close together, and the longest run of at least four is proposed as a rectangle.

That is a heuristic, written down here so it is clear what it is and is not. It has not been measured against a corpus of real scans, so the hit rate is an honest unknown. It will miss a white or black software timestamp entirely, because neither passes the colour test. It will occasionally offer a rectangle over a string of fairy lights or a row of reflections on water. And if more than a twelfth of the searched band passes the colour test at all — a sunset, a bonfire, a stadium at night — it concludes that colour is not distinguishing anything here and proposes nothing rather than guessing.

Why is the date orange and not white?

Because it was exposed onto the film rather than printed onto the paper. The imprinting unit sat behind the film gate with a small red or amber LED, and colour negative film records that as a dense orange-yellow when it comes back as a positive. A white date on a photograph almost always means a digital camera or a print shop drew it afterwards, which is a different problem with an easier answer.

Is there any way to get the date back if I change my mind?

Keep the original scan. Nothing here alters the file you opened. What you save is written fresh, and in that copy the digits are gone for good, because the pixels holding them were overwritten. If the date matters — and on a family scan it very often does — put it in the file name or the folder, which is where a date on a scan is actually useful.

My imprint is over somebody's face.

Then take the crop instead. A fill has no idea what a face is and will return something plausible and wrong, and on a photograph of a person that is the worst possible failure. Either live with the date, or crop the frame so the corner is outside it and accept the composition you get.

The scan is grainy and the fill came out smooth.

That is the marching fill doing exactly what it does: averaging nearby intact pixels, which averages the grain away with them. Switch to the patch search, which copies real grain from elsewhere in the same scan rather than computing a mean. On a heavily grained scan that difference is the difference between a patch you can find and one you cannot.