How Photoshopped Is That Picture?

Freddybear writes "Digital forensics experts at Dartmouth have developed software that can analyze digital photos to rate how drastically they have been altered by digital editing techniques. 'The Dartmouth research, said Seth Matlins, a former talent agent and marketing executive, could be "hugely important" as a tool for objectively measuring the degree to which photos have been altered.'"

Comments

Re:It Can Tell by the Pixels

[kelemvor4]

The whole term is dumb. You don't photoshop something in GIMP, for instance... just like you don't xerox on a Cannon.

http://dictionary.reference.com/browse/xerox

noun 2. ( sometimes lowercase ) a copy made on a xerographic copying machine.
verb (used with object), verb (used without object) 3. ( sometimes lowercase ) to print or reproduce by xerography.

See also: http://dictionary.reference.com/browse/photoshop — vb , -shops , -shopping , -shopped ( tr ) to alter (a digital photograph or other image), using an image editing application, especially Adobe Photoshop
Notice it says especially, not exclusively.


My point is that you most certainly do photoshop something in G.I.M.P..

Re:Having a little experience here

[squidflakes]

Most DSLRs have "HDR" capabilities. I know that on Nikon it is called bracketing and you can take three or four images with up to a 2 EV increase or decrease on on either side. That still only gives the image out of your camera 10 - 12 EV of range, which bumps right up against what most monitors are capable of displaying.

Re:Celebrity culture...

[hankwang]

And even before photos are loaded up in Photoshop the celebrity has already been loaded up with a pound of makeup, sat under carefully positioned lights and been photographed by a professional.

Indeed. Dove Evolution video clip

Re:Revert?

[dgatwood]

Creating a Photoshop job that will fool even a sophisticated human eye is easy. Creating one that will fool an algorithm is very, very hard. The modification detection algorithms I remember hearing about start by taking a series of test images with the specific camera that was used to take the picture. In order to beat them, among other things, you'd need to:

• ...mathematically compute the probability of noise for each subpixel and adjust your noise so that the distance of each pixel from the mean of nearby pixels in areas of low contrast is close enough to what would be expected for that particular spot on that particular CCD at a particular physical temperature, and so that the noise level is consistent with the expected noise for a single physical temperature value across the entire image.

  Alternatively, if a particular camera gets hot spots on parts of the chip when shooting lots of pictures in a row, the noise level might need to be a very complex gradient with the hot spots in particular places on the chip.

• ...know where every dead subpixel is on that camera so that you can mathematically compute the correct color channel value for each dead subpixel based on its neighbors in the same way that the camera does.
• ...read the EXIF data to determine which pixels the camera mapped out because of dust, if that particular camera does that, and compute their values programmatically in a similar fashion.
• ...precisely reproduce the chromatic aberrations of the lens at every point in the image.
• ...precisely reproduce the subtle variations in tint at every high contrast edge caused by the relative positions of the subpixels at that particular point on the CCD.

And so on. It's not a case of artistic training. It's a case of spending months modeling a single camera in MatLab. Not a single model. A single camera.