Secure Computing: Sec-C

There are a few coincidental truisms I have come across over the years. If the restaurant has "Golden" in the name, then it means inexpensive, low quality buffet. Golden Corral, Golden Wok, Golden Emperor, Golden Palace Seafood, ...

If anything has the title of "World's Best" or "World's Greatest" and it isn't in the Guinness Book of World Records or awarded by some judging society, then it is certain to be neither the best nor greatest.

And if something says "Secret", then there's probably something obvious. Years ago I moved into a new apartment and went kitchen shopping with my mother. I pulled out a cookie sheet and said, "Baker's Secret? I wonder what the secret is?" My mother whispered to me, "Real baker's don't use it."

So you can imaging my skepticism when, over the years, I keep seeing commercials, print ads, and mall outlets for a company called "Victoria's Secret". What's the secret? Photoshop.

Disasters

One of my favorite web sites is Photoshop Disasters (PD). I'm always amazed when people can just look at a picture and identify subtle abnormalities. (My friend Cynthia Baron does this all the time. Really impressive.) Most of the pictures featured on PD are subtle, but when pointed out they become obvious. I mean, yeah -- a baby with seven fingers is clearly photoshopped. However, unless someone pointed it out to me, I probably wouldn't have noticed. When looking at the PD images, I only see about half of them right off the bat. (Dude! Where's your legs?) The other images may take me a minute or two.

I've been using many of the pictures on PD as a testbed. The PD community usually points out one or two things that are visually obvious; a distorted limb, wrong shadows, incorrect perspective, etc. However, artists rarely just do one or two things to these photos. Using my analysis tools, I've been trying to find "what else" was changed...

Remodeling

One of the recent PD entries features a photo from Victoria's Secret.



The PD community quickly identified a few obvious items:

• Her bellybutton looks like a black hole to hell.
  
• For someone so thin, she is missing bones... no ribs nor hips.
  
• While her hands are casting a shadow onto her body, the blue ribbon casts no shadow.
  
• Her legs are different shapes. (I'm willing to say that this could be due to posture, but they pointed it out.)

Visually, I saw a few other items:

• The bellybutton does not appear moved or distorted -- it is just darker. And it lies right where I would expect the blue ribbon's shadow to be. So the artist must have lightened her belly, but forgot to lighten the bellybutton.
  
• The shadows above the bikini top are dark and crisp, but the shadow from the hands are faded and light. In fact, her hands have shadows, but her finger-shadows are missing. Her belly was seriously lightened.

But I rarely, trust my eyes... Two of the algorithms I use identified much more...

That's Dense!

The Color Density Distribution (CDD) is a fun algorithm. Basically, it tries to find color variations that have been distributed around a centralized color. With computer graphics, texture and coloring is done by selecting a single color and then permuting the normal to the surface (when rendering) or distributing colors around the single color. With CDD, these variations will recombine back onto the central color.

In contrast, real life is not uniformly colored. A gray desk actually has many different gray colors -- and then the lighting creates more variations. There is no single color for clustering. Items like walls, sky, and skin will appear blotchy rather than having a solid color. If you see a solid color, then it likely indicates manipulation. (There are a bunch of caveats here, but this description is good enough.)

Here's the CDD of the Victoria's Secret picture:

• Her upper chest, arm, and most of her face look blotchy and mottled -- realistic. However, there is a solid red patch on her face: from her eyes to cheeks, and crossing over the bridge of her nose. That is not natural; the artist modified her face.
  
  
• Her nose consists of multiple colors (realistic), but her lips are a solid red (likely recolored).
  
  
• Her belly is a large, solid color. And rather than getting very blotchy at the edges, it smoothly fades. This is manipulation; her belly has been recolored. In the process of remodeling her belly, they removed shadows and bones. Because the transition is so smooth, I cannot rule out the liquify tool (for making fat people look thin), but the association with liquify is not conclusive; it could just be some serious recoloring.
  
  
• The blue ribbons -- the one missing the shadow and the two stripes on the bikini bottom -- are not mottled or blotchy; they are solid. They have been recolored. We could even speculate that they may have been added. However, since the skin directly under the blue ribbon is blotchy, the bow probably did exist and it was just the belly and blue coloring that was changed.
  
  
• Most of the natural blues in the image (water and fence) become blue under CDD. However, the blue between her legs becomes black and green. This means that those colors are actually not the same colors found on the fence. I really believe that the original photo had no space between her thighs, and the artist created the space between her legs.

There is one other abnormal area: the background. The sky becomes white and the hills become green -- no splotches. However, the original picture is blurry. A real photo with a blurry area will appear like a Gaussian blur when the image is resized. And Gaussian distributions are centered around a specific value. So these could be natural.

What a DoG!

One algorithm that I'm beginning to look at is the Difference of Gaussians (DoG). DoG is usually used for identifying edges (using the absolute difference), or for sharpening images.

What I'm finding is that DoG edges are usually present in all color channels. The edges should appear grayish or white-ish. There are some occasions where they will appear a little colored, but that's a less common instance. However, with CG or manipulated areas, DoG edges may be totally absent from some color channels.



In this case, the circle patterns on the bikini are far too colorful. Each circle is restricted to one or two color channel. The circles have either been enhanced or added. I guess this explains why Isha wrote in one of the PD comments:

For the clothing, I think they shoot and then paint on the pattern so that the potential consumer can "see" it better. It seems to be a common thing for catalogs when you have the same style in multiple colors/pattern or if the original photograph doesn't show it in its vibrant "true" (hah) color.

Socialism

Of course, there is a social impact here. All of those impressionable teenage girls who barf up their breakfasts in order to reach some ideal thinness that matches these supermodels... Well, even the supermodels don't look like that. Rather than ruining your health, go buy Photoshop.

Update: Victoria's Secret revised the image. The updated picture is evaluated at Still A Secret.