Sharpen an image

What sharpening actually is

It increases the contrast at edges that are already in the picture. It does not reconstruct anything. A photo that is genuinely out of focus, or blurred by camera shake, has lost that information at the sensor — and no amount of sharpening brings it back, it only makes the blur more contrasty.

Where it works beautifully is on the very common case of a photo that is slightly soft: shot at a small aperture, resized down, saved as a JPEG a few times, or scanned. All of those attenuate fine detail without destroying it, and that is exactly what an unsharp mask restores.

The three controls, and which one matters most

Strength is how much. Past about 2 you start seeing halos — a bright fringe on one side of every dark edge — which is the unmistakable signature of an over-sharpened image.

Radius is how wide an edge counts as an edge. Around 1 pixel crisps fine detail, which is usually what you want. Larger values stop sharpening and start adding local contrast, which can look striking and can look artificial.

The threshold is the one that separates a professional result from an amateur one, and it is the one nobody adjusts. Sharpening every pixel equally amplifies sensor noise in the areas that should be smoothest — a clear sky, skin, a sheet of paper. The threshold says "only sharpen where the contrast is already above this", so flat areas are left alone.

The threshold is measured, not guessed

A single default cannot suit both a studio photograph and a phone snap taken indoors, because the right threshold depends on how noisy the image is. So by default this measures it.

The noise floor is estimated from the median absolute deviation of the image's Laplacian — a standard robust estimator, which works because the median ignores the small proportion of pixels that are genuine edges. Checked against synthetic noise of known strength it reads within about 30%, and on an ordinary photograph it reports a little under 2 grey levels.

The threshold is then set at two and a half times that. The effect is measurable: on an image with a noise level of 9, a fixed threshold of 8 let flat-area noise rise from 6.0 to 14.4, while the measured threshold of 22 held it at 7.2. On a clean image it chooses 3 and sharpens freely. The result states the figure it used, and you can override it.

Denoise first if you need to

On a genuinely noisy image, cleaning up before sharpening beats raising the threshold, because the threshold protects flat areas while a denoise actually removes the grain. The two work well together and the order here is fixed — denoise, then sharpen — because sharpening first would amplify the noise you were about to remove.

Up to twenty at once

Drop in a set and they come back as a zip, each one thresholded from its own noise rather than from a shared assumption. The reported threshold becomes a range so you can see how much they varied.

Save as PNG if you are going to edit again

Sharpening is not reversible and JPEG compression is not either, so sharpening a JPEG and re-saving as JPEG applies two lossy steps to the same file. If this is a step in a longer process rather than the final output, PNG costs you nothing but space.