Who Stole My Pixels?

[PierreVandevenne]

Which is it, Warren?

Just aliasing. You aren't sampling his messages at the correct frequency. You could try interpolating an interpretation, unless you have reasons to believe one of the statements is an artifact. ;-)

[ejmartin]

That depends on the subject, and on the demosaicing technology used, but it is usually better than people think, or want us to believe. I've tried to determine the real loss of resolution, and it can be as little as 6.4% for luminance only. Here is a simple experiment I documented 6 years ago. This was before I got hold of real (deconvolution) sharpening routines instead of simple edge contrast enhancements like USM.

Cheers,
Bart

Nice demonstration.  What algorithm did you use for the demosaic?

BTW, Bart, is there an index page on your site that links to all your articles?  I can't find it by backing out of the above URL (or the one on aliasing).

[ejmartin]

Just aliasing. You aren't sampling his messages at the correct frequency. You could try interpolating an interpretation, unless you have reasons to believe one of the statements is an artifact. ;-)

         

[Bart_van_der_Wolf]

Nice demonstration.  What algorithm did you use for the demosaic?

Hi Emil,

I used DMMD's "Pictura" beta software, but it's no longer available at their site (i.e. I cannot find it there anymore). It used a so-called "Adaptively Quadratic (AQua) Image Interpolation" method which produced good demosaicing results.

BTW, Bart, is there an index page on your site that links to all your articles?  I can't find it by backing out of the above URL (or the one on aliasing).

No, there is no index, and backing out is blocked. I wrote most of those short articles to avoid having to answer the same questions over and over again in Usenet discussions (they are still useful for that purpose today it seems).

There are additionally only some articles about scanning (mainly oriented at the Minolta SE 5400), but I never finished the entire analysis because the model was succeeded by a newer one, and Minolta was abandoning that market segment, so I didn't feel motivated to 'recommend' that scanner anymore. Service was also becoming problematic, and 35mm film scanning was being replaced by 16+ MPixel cameras with better enlargement potential. I still use the scanner for my old films, and Vuescan as it's driver because there is no support for modern operating systems from the manufacturer.

Cheers,
Bart

[Theresa]

If it bothers you so much get a Sigma.  Although it lies even more about its non-Bayer sensor.  It claims three times the resolution which it doesn't even get close to.  All cameras have to work with GRBG except for the Sigma which is RBG.  Since it is universal it should just be considered a given and take into account that the results are so good.  They would be more accurate if they gave the interpolated values but that's modern marketing.  You can make a print several feet in length with todays sensors so they are doing something right.  If you are really adventurous you could have your anti-AA filter removed which would allow you to get the most from the pixels that are there, as long as nothing causes moire.

[JBerardi]

We have all been conned.

Trolled is more like it.

[dwdallam]

We could turn it into a productive effort though. Is Foveon better than Bayer?

[WarrenMars]

Seriously. Did your pictures lose 3/4 of their appeal after you found this out?

Amazingly Yes! I was so disappointed that I had to get rid of 3/4 of them.  

More seriously, I'm surprised so many of you think interpolation gives you back 3 times the information you actually have! You need 3 colours to make a real pixel, there's just no way around that fact. The extra green is simply a waste of time put there to make the grid a square. Thus it is that in a Bayer array you need 4 pixels to make 1. Your true resolution IS 1/4 the advertised.

You can delude yourself that sophisticated algorithms put back what wasn't there in the first place, but anyone who has studied noise theory should know that is not possible. All you can do is GUESS, which is what interpolation does. Sure, your guesses might be correct a good percentage of the time, but that's not the same thing as having the ACTUAL data. You wanna shoot motor cars and chick's faces? Interpolation is for you! You wanna shoot foliage and landscapes? Better get some more pixels!

Foveon's multi-chromatic sensor sounds good in theory, but a look at Sigma's output will quickly convince you that it doesn't produce in reality. If the answer is in that direction they're a long way from it. I'll tell you something that WILL improve the situation: A three colour filter!

The conventional filter is 4 squares to make a square, wasting 1/4 of the area. But we don't HAVE to have square sub-pixels! Let's use 3 rectangles of width 1/3 the height. 3 rectangles to make a square! This will improve the resolution by 1/3. Not huge, I'll admit, but a useful improvement!

Go to it Manufacturers!

[bradleygibson]

Or, just consider the camera you already have to posess 1/4 its stated pixel count.  Voila!  Fully sampled 'macropixels' with 2x oversampling in green!

(And way better S/N performance than the 'skinny sensel' solution).

QED.

[PierreVandevenne]

is simply a waste of time put there to make the grid a square. Thus it is that in a Bayer array you need 4 pixels to make 1. Your true resolution IS 1/4 the advertised.

If you were capturing an image at a single (or a small range of) wavelength(s) matching the transmission curves of the respective filters in the array, roughly yes, you'd have 1/2 in G, 1/4 in R and 1/4 in B. However, that usually isn't the type of images photographers captures.

And believe it or not, researchers have been looking for better sensors for a while

http://johncompton.pluggedin.kodak.com/def...asp?item=624876

Incidentally, note that their main purpose isn't what you think it should be.

You'll be, I am sure, "SHOCKED" to discover that not only did the CFA steal your pixels, it also refused entry to your photons!

[Jonathan Wienke]

More seriously, I'm surprised so many of you think interpolation gives you back 3 times the information you actually have! You need 3 colours to make a real pixel, there's just no way around that fact. The extra green is simply a waste of time put there to make the grid a square. Thus it is that in a Bayer array you need 4 pixels to make 1. Your true resolution IS 1/4 the advertised.

Welcome to 2003, Warren. While Bayer matrix does not sample every color at every pixel location, but it does sample something at every pixel location. So with a properly-designed AA filter to ensure there is some degree of correlation between every pixel and its immediate neighbors, one can accurately "guess" (interpolate) average luminance for every pixel in the image, at the expense of color accuracy in fine details. Green appears twice in the matrix because human vision is more sensitive to green than red or blue, and color errors in the green channel are more visible than the red or blue channels. It's not "simply a waste of time put there to make the grid a square", it's a carefully considered engineering compromise that allocates the most real data to the wavelength range that needs it most, based on the perceptual response of human vision.

The net result is much better than you claim; if you compare Bayer images to Foveon images (which do sample all three color channels at each pixel location) you will find that Foveon images are better pixel-for-pixel than Bayer images. After the proper sharpening for each image (deconvolution-based, NOT simple USM), the Foveon is only about 1/3 better; a 4MP Foveon image need between 5 and 6 Bayer MP to match overall resolution and detail. If you were correct, the Bayer camera would need 16MP to match a 4MP Foveon image, and if you actually compare images, this is far from the case.

[Jeremy Payne]

...

Pass whatever you are smoking down this way ...

[JBerardi]

Pass whatever you are smoking down this way ...

I dunno man, I think he got some bad stuff. Paranoia and all...

In all seriousness, we can argue about the math, the physics and the technology all day long. Ultimately though, it doesn't matter if my camera is really the 10mp it claims or not. That's just nomenclature. What matters is how the 12x18 hanging on my wall right now looks. As a matter of fact, it looks great. It's kinda like saying that a car isn't really a V6 because all six cylinders don't fire at the same time. It's an arbitrary distinction. It's the real-world performance that matters.

Besides, even if we somehow forced manufacturers to state the "correct" number of megapickles, it wouldn't do anything but confuse people. Even if the numbers aren't technically correct, everyone knows what kind of resolution they can expect from a 10mp camera, or a 24mp camera.

[pegelli]

Wow: I'm amazed nobody started complaining about this blurring filter these villain manufacturers put in front of the sensor to drop the achievable resolution even lower    

On the other hand the jpg's and tif's I get from my camera meet the pixel count they claim so I really don't get what's the problem  

OK, let's get serious. I like this thread as some of the technical explanations about the subject increased my understanding of what's really happening. Thanks for those  

[tokengirl]

It was a one-eyed, one-horned, flying purple pixel eater...

[SandroD]

that test he did seems to me absolutely pointless...

[PeterAit]

No doubt many readers know this already but I was truly SHOCKED to discover that each supposed pixel in my camera contains only the data for one colour! This means that the true resolution of my camera is only a QUARTER of its advertised value! My 10MP SLR is really only 2.5MP!

I have been aware of interpolation and the demosaicing step for years, but I had always assumed that this was to make small realignments to centre the 4 elements of the Bayer mosaic within the pixel. It never occurred to me that it was being used to invent colour data for three quarters of the image!!! It was only when I was stepping through the main data structure at runtime in Dave Coffin's DCRAW code and wondering why only one quarter of it was filled in that the penny finally dropped. Shocked I tell you!  

Ken Rockwell is right. The manufacturers are LYING! Interpolated pixels are not real pixels. Made up data is not real data. We have all been conned. Sure, the interpolation used these days is very good! In areas without sudden transitions the invented detail is accurate, but in areas of fine detail: Ouch! It isn't!  

There are laws about misrepresenting the specs of one's products. A minor inflation of one's resolution might be acceptable, but 400%? Surely that's illegal, not to mention immoral.

Maybe you should alert the UN, or Homeland Security, or the FBI. This is REALLY IMPORTANT!

Sensors don't have pixels, images have pixels.

[stamper]

Maybe you should alert the UN, or Homeland Security, or the FBI. This is REALLY IMPORTANT!

Sensors don't have pixels, images have pixels.

Maybe he was pixelated when he wrote it

[Eric Myrvaagnes]

You guys got me worried about Pixel Theft, so I dropped everything and started counting the pixels on my Canon 5D's sensor. Unfortuantely, I fell asleep less than halfway through the second row, so I don't know whether any of mine are missing.

I guess I'll have to get a Pixel-Peeping Watch Dog to guard any remaining pixels.         

[WarrenMars]

It was a one-eyed, one-horned, flying purple pixel eater... wacko.gif

Mos def the funniest line so far!  

And believe it or not, researchers have been looking for better sensors for a while

http://johncompton.pluggedin.kodak.com/def...asp?item=624876

Interesting to look at Kodak researchers work. No question as to why Kodak are going out the back door.  So they're gonna do the old "get the missing colour by subtracting the other 2 from the total" trick, as used in DVDs. Not a bad scheme I guess, but why include the 3rd colour at all if you're gonna go that???

Anyway, it's an extra level of complexity that we don't need, although it will give the same resolution improvement as my previous suggestion.

It may be that rectangular pixels are tricky to make micro lenses for, (perhaps someone else might have a handle on this?), so how about this for a better array?



Each triangle is an image pixel. You get 2x the image pixels that you have in reality. This is cheating folks, but it's NOT interpolation! What's more, it's dead easy! Just use the values of the 3 corner pixels. SIMPLE AND ACCURATE!

Ok, folks, tell me it's been done already...