Actual number of light-collecting sites in GH1?

G

guyburns

Active member
I just want to make sure my limited understanding of digital light collectors is on the right track. Let's assume that the GH1 has 4000 x 3000 pixels. Does that mean, assuming a 2x2 (RGGB) Bayesian pattern for its light sensor, that the sensor actually consists of an 8000 x 6000 array of light collectors?

Therefore, during processing of this information for a stills image, each 2x2 array is converted into a single pixel.
 
No, because Bayer doesn't work in a 2x2 pattern. Demosaic'ing algorithms use many more sites and average the image across many sites.

Strictly put, the GH1 has 14 million light-collecting sites. Those get de-bayered into a collective 1920x1080. But in video mode we don't know how many of those 14 million sites are actually used.
 
First of all, I'd better correct the wrong terminology in my post. I should have said "Bayer pattern" not Bayesian pattern. Bayer and Bayes are two different people, centuries apart -- but they come together in digital sensors because the mathematics of Mr Bayes can be used in the demosaicking of a Bayer pattern.

I should have made clear in my post that I am talking about "stills", so as to not complicate matters by introducing how the GH1 downsamples its sensor data for video.

Barry, if you are correct, the GH1 has the same number of light-collecting sites as the advertised megapixels. For a single-sensor camera, is that an industry standard across all cameras? If a digital camera is advertised as 20 MPixels (say 5000 x 4000), the sensor has 20,000,000 light-collecting sites?

QUES: Are you sure that the GH1 only has 14 million light-collecting sites? How do you know it doesn't use a 2x2 Bayer pattern and actually has 14 million red, 14 million blue, and 28 million green? If they did, I suppose they might be tempted to use the higher numbers in their advertising. On the other hand, I think of LCD screens which are advertised as, say, 1920 x 1080, whereas they actually have three times that number of light emitters, but manufacturers never use those higher numbers in their advertising.

I was surprised when I found out the extent of "throwing away" that the GH1 does in extracting 1080p video from its sensor, now I find out that it is effectively upsampling in a serious way to get it's advertised "still" pixels. Not just the GH1 -- all digital cameras.

I've got no problem with that, it's just that I thought a digital camera would use a "cleaner" approach. But I guess because of technological constraints it isn't possible.

In rough outline, is this how it works? Assume the simplest 2x2 Bayer pattern is being used:

R G
G B​

and the camera has 4000 horizontal pixels and 3000 vertical. Remember, this is just a simplification to aid in understanding:

1. Starting from the top left, the light-collecting sites in the first two rows would be:

R G R G R G R G... (4000 in total)
G B G B G B G B... (4000 in total)

and so on for the next 2998 rows.

2. The top left pixel in the final image is obtained by demosaicing the first 2x2 Bayer pattern; then you move to the second Bayer pattern; and so on across the first two rows.

3. Then you demosaic rows 3 & 4 to obtain the second row of pixels in the final image.​

Proceeding as above would result in a final image of 2000 x 1500 i.e. 3 Megapixels, but the GH1 does better than that. According to:


the GH1 can resolve 2300 lines in the vertical direction, reasonably close to the 3000 actual light-collecting sites in the vertical direction. Yet it does so under the constraints of its Bayer pattern, which if implemented as simply as I have described would only be capable of resolving 750 vertical lines (3000/2 = 1500 "real" vertical pixels after the Bayer pattern was taken into consideration, divided again by 2 because of Shannon's sampling theorem).

Obviously some advanced signal processing is going on, and it leads me to wonder just how much of the image we see from digital cameras is actually real. To give an example from a different field, if you've ever listened to a VoIP phone call using the G.729 codec, none of the voice is real. The whole lot is computer generated from scratch, based on the characteristics of the speech extracted in real time. What you hear is a synthetic voice that sounds remarkably like the original. It's not sampled, it's synthesised.

The reason I posted this question is because I couldn't find out how many light-gathering sites there are in the GH1. I'll take Barry's figure as correct, unless someone else cares to dispute 14 MB.

If anyone else is interested in a deeper understanding of demosaicking, and you don't mind getting technical, here's a good start:


It is clear, after reading that article, that demosaicking is still very much a work in progress, and the flaws when using some methods to demosaick may explain why colour fringing is more obvious in the video from some DSLR models than in others (for example, Figure 6 in http://www.imageval.com/public/Papers/DigitalCameraSimulation_092908.pdf).

If anyone knows of any papers that specifically reference the innards of a movie-making DSLR, please post.
 
You must log in or register to reply here.
Back
Top