Slashdot Mirror
HP Restores Creased Photos With Flatbed Scanners
An anonymous reader writes "Scientists at HP have developed a technique to detect creases in photographs using standard, unmodified flatbed scanners. Once correctly scanned into a computer, software can determine where the photograph's defect is, and artificially correct it to remove any trace of a crease or fold. The result is a spotless JPEG scan from a creased photo, with absolutely no modified hardware and no technical know-how required on the part of the user." They're using multiple light sources to do this, in a way that reminds me of last year's description of 3D image creation using an ordinary digital camera.
A fold-less centerfold :)
Once I was a four stone apology. Now I am two separate gorillas.
Won't this ruin my collection of photographs of creased paper?
In the article it says that they use an unmodified scanner, but later on they claim to control the lights of the scanner individually... how is this not modifing the hardware?
I was hoping they were using that 3D information to do something interesting to actually restore the image. They're not.
They're basically using rudimentary 3D information that they can get out of the scanner to determine that a crease exists. They then remove it with a simple infill algorithm, which is as basic as it gets (although it often works ok), and which you can find in most image editing software. It's no coincidence that the example image they use has a crease going over mostly similarly colored and low-detail areas.
So what they're doing is not an improvement to restoration, it's just an improvement to defect detection. Basically, it saves you having to tell the software where the defect to be fixed is, the fixing is the same quality as it's always been.
--Joakim Ziegler
The rudimentary 3D info can be used for improving all sorts of scans.
How about...
- Flattening a scan of a book (by the spine)
- Focusing an area that's raised (products like Focus magic assume a section is all out of focus at the same level, whereas a map of the amount of lost focus is possible here).
- Using the above, scanning non-flat items.
- Scanning nearly-flat 3d surfaces.
Add a lens that can vary focus (based on the light differential) and you'd have a good 3D scanner for one side of a mostly-flat item, and a flatbed scanner that wouldn't lose focus on slightly-raised papers.
What we really need is a copy machine/scanner that can detect the valley formed by the spine of a book being copied and automatically correct for it. That would be worth it.
excitingthingstodo.blogspot.com
Multiple light sources offer some interesting options. A few years ago, someone modified a digital camera (I think a Canon PowerShot) to have four flash sources instead of the usual one. The camera would take four pictures in quick succession, one with each flash. This allowed better edge detection.
It was useful for applications like taking a picture of complex, dirty machinery (as under a car hood) and locating the edges, even where everything was roughly the same shade. It also helped when photographing very shiny objects, where the reflection from the flash was a problem. With each reflection from each flash unit in a different place, all reflections could be removed.
It was too specialized to become mainstream, though. That seems to be the fate of 3D from 2D systems. Good ones have been built, but most have been either discontinued or turned into very expensive products for specialized use.
From the examples shown in the .PDF [hp.com] it seems that it is once again a case of a quick fix that only works on low-res and low detail photos, preferably in single color.
That doesn't seem like a terribly bad thing to me. If you were a professional looking for extremely high-quality results, then yes, you're going to want to spend a lot of time screwing around with things manually on each photo. Even if it's a largely automatic procedure, you'll probably still want to tweak the parameters a little for each photo, including things like brightness, contrast, and hue.
However, there's another real-world application for this sort of thing: someone like my grandmother scanning lots of old pictures that may have been folded, crumpled, or otherwise damaged. Even if it's not giving the highest quality results, if the results are at all better than not processing the photo, then it's probably fine. Without automatic quick fixes, people might either scan it and leave the damage, or decide not to scan it at all. Giving even barely passable results is an improvement.