[drhagen]

I'm still skeptical. He blurred it and then cropped it. The sharp edge from the cropping is going to mess up the deconvolution optimization by causing any and all deconvolution to strongly reduce sharpness along that edge. If I'm right, the Samsung would restore both halves equally well if the blur was applied after the cropping rather than before.

I still want to see this whole thing done without Gaussian blur, which is readily reversible by even primitive algorithms.

[redox99]

> which is readily reversible by even primitive algorithms.

Not true unless you know beforehand the exact kernel. Besides, there's 0 reason why a camera software be programmed to, or attempt to revert Gaussian blur.

And finally, he downsampled the image, and you can't recover from that.

[geoduck14]

Isn't a gaussian blur a pretty standard blur done by a poorly focused camera? Perhaps Samsung went to great lengths to reverse a Gaussian blur because it happens all the time?

[jrockway]

Gaussian blur and the blur from out-of-focus lenses are a different kind of blur: https://en.wikipedia.org/wiki/Defocus_aberration https://en.wikipedia.org/wiki/Gaussian_blur

[Jasper_]

No, a real camera would give you a box blur, because the out-of-focus light rays are distributed uniformly across the sensor.

[toxik]

Gaussian blur is indeed super common in nature, any sum of noise will be Gaussian in the limit.

[sorenjan]

Perfect deconvolution of a gaussian blur is only possible in theory, or in ideal cases where you control the signal end to end, like a programming assignment. In reality there's plenty of sources of errors, like discretization noise, lens blur, sensor noise, and SNR is instrumental to the result. I've never seen a blurry image get reconstructed using deconvolution without artifacts.

Then there's the fact that a photo enhancement program has no reason to even try deconvolving gaussian blur, they should use a more appropriate kernel that better describes out of focus blur.

His original blurred moon image is available in his first post. Without even taking the detour of photographing it from a monitor, try using his image and any deconvolution software you want, and try getting anywhere near his result.