Adobe deblur technology - amazing

[gumboot]

If I'm reading between the lines correctly, this trick is very effective with motion blur, particularly if the motion is more-or-less constant during the exposure. It won't help much with defocus or jitter.

It reads to me more like it will be effective with jitter (camera shake) and not defocus or motion-blur.

 

[gumboot]

I've always wondered if this violates some principle of information theory. For that matter, does information theory give some upper limit on how badly something can be blurred and still recovered?

It really depends on what you're attempting to extract. The sort of thing shown commonly on CSI-type programmes is impossible. But as Darat points out there is already technology that can pull letters and numbers from a blurred car numberplate. This is a controlled context, however, and the computer knows exactly what it's looking for.

In "real world" the computer has no way of knowing what it's looking at, and therefore what it "should" be seeing. Any attempt to "fill in the gaps" would be nothing but a blind guess.

It also, of course, depends on the degree of blurring. If you're talking about a correctly focused picture with movement blurring you might be able to extract detail because it's still there in the picture, it's just moving. That's what this new plugin does; it detects the movement of the camera and then reverses it.

By contrast a picture that's actually out of focus doesn't contain the necessary information. The more out of focus the picture is, the less information can be provided, even if the picture can be "sharpened" using edge-finding filters. It will be "sharp" but it still won't have a lot of detail.

Of course all of that's a whole separate story from the classic CSI scenario where they're not addressing blurring, but actually artificially increasing the resolution of the image. That can be done to a very, very limited degree on things that are familiar such as text. I would imagine you would need at least three or four pixel height text for it to work however. If, for example, an entire page of a newspaper occupied a single pixel, there's no hope of resolving what's on it.

It can't be achieved with non-text subjects because once again the computer has no way of determining what it's actually looking at. Is that red patch of pixels a sheet of flowing, smooth silk, or a rough textured piece of canvas?


ETA.

One last crucial thing to bear in mind is that there's obviously an enormous gap between "correcting blurring to show a picture that looks sharp" and "correcting blurring to show a picture of what's actually there". It's fine, for a general photograph, to have a computer fill in the blanks with information of its own determination. In regards to police investigations and so forth, they need to determine what was actually there, and that's a much trickier proposal.

It would be very interesting for example to compare the "deblurred" plaza photo in the OP with a sharp photo actually taken in the location at that same moment. There's an enormous amount of artifacts in the sharpened photo and I wonder how much of what it depicts is actually accurate.

 

[This is The End]

The sort of thing shown commonly on CSI-type programmes is impossible.

I always try to point out that while it is impossible to increase the resolution of a single image, it is quite possible to increase the resolution if that image is one of several.

For instance, if the image is a frame in a video. All of the frames of the video have more information in total than they each do individually.

So if you have video of a car driving by, you may not be able to read the plate # in any specific frame. Yet all of the frames that the plate # is visible in may have enough information in total to compile a very good resolution image.

There are a few complex routines that computers use do that, but a very simple one is to just center & align the target (in this case the license plate) in every available frame, and then layer the frames while averaging out irregularities.