Fix an out-of-focus photo: what is measured
To fix an out-of-focus photo the page first needs the shape and size of the blur. A real lens out of focus turns every point into a small uniform disc; general softness is closer to a Gaussian. The disc leaves rings in the spectrum and the Gaussian does not, so the page looks for the rings first and falls back to the width of the strongest edges.
| Kernel | Estimated from | Range here |
|---|---|---|
| Lens disc | First zero ring of the radially averaged spectrum | 0.5–12 px radius |
| Gaussian | 10–90 % width of the strongest straight edges | 0.5–12 px sigma |
Where it stops
Past a radius of about twelve pixels at the working size the solver averages more than it recovers, and the triage on the unblur page calls the photo too far gone. Below that, the engine report publishes the error before and after on discs of 1.5 to 8 px and Gaussians of 1 to 4 px; the build fails if the error stops falling by at least 30 %.
Related on this site
- Not sure it is focus rather than shake? The triage page measures which kind of blur it is: unblur an image.
- If the smear has one direction, it is camera shake, and the motion solver fits it better: remove motion blur.
- The estimator's accuracy and the solver's error, measured against known plates, are in the engine report.
Questions: fix an out-of-focus photo
Can an out-of-focus photo really be fixed?
Partly, when the misfocus is mild. A lens focused on the wrong plane spreads each point into a disc; the spread can be divided out where the lens still passed some of the detail. Where the disc is large, the finest detail was averaged away when the picture was taken and is not in the file, so no method can bring it back. The page measures the disc and says which case you have.
How does it know how out of focus the photo is?
Two ways. A lens disc leaves rings of near-zero energy in the picture's spectrum, and the first ring's position gives the disc radius directly. When there are no rings — general softness rather than a lens disc — the width of the strongest straight edges gives a Gaussian blur estimate instead. Both are shown with a confidence figure, and both are checked against plates blurred by known amounts on the engine report.
Why does the result show ripples near strong edges?
That is ringing: the division overshoots where the lens passed almost nothing. The noise term is the floor under the division; raising it trades recovered detail for calm edges. The Hold ringing under 3 cv button measures the ripple on your own loupe patch and raises the noise term until it falls under three code values, then says what it changed.
Is this the same as sharpening?
No. Sharpening steepens the edges already there. This inverts a measured blur — a different operation that can return some detail sharpening cannot, and only when the blur model fits. The saved file ends in -deblurred when the solver ran.
Is the photo uploaded?
No. The measurement and the solver run in a worker in this tab on your own processor. There is no upload endpoint on this site.
