Nokia 808 - reading between the array lines

Nokia seems to have created quite a storm over its new 808 camera. Personally I'm very interested in it because it seems to have kept Symbian as the OS (gosh won't win many friends saying things like that in this iPhone and Win7 phone climate).

But I still happen to think that the OS is overblown and its the functionality of the phone which is paramount.

Anyway, this post has its focus on the Camera in the phone ... DPReview has a quick intro to the phone over here, which I suggest you go and have a look at.

Basically I'm wondering why a 41MegaPixel phone camera would be of interest? Particularly when it normally saves images as something significantly smaller (like 5MP). Heck I already have problems with being able to use my phone to send an MMS to people because the 5MP exceeds the capacity of the automated downscaling. Not to mention the problems I have with family sending me 10MP images in email as attachments (and choking my inbox or them getting bounces).

Anyway the DPReviewpage puts up an image as a sample which is actually quite interesting. Its called "quiet moment" and is (according to the EXIF) taken at 800ISO. I've taken that for examination and (to avoid any copyright annoyance) have added to it below to show where I'm interested in examining.

Firstly, lets get to some basics, Nokia is claiming that the camera is 41Megapixels, which would of course be wild. However its worth reminding that Megapixels and resolution don't just go hand in hand. Its a square relationship.

So this camera produces 41Megapixels, which is 7728 x 5368 pixels. Now the claim is that it

• produces very clean output when sampled down (using pixel binning)
• and allows for great lossless digital zoom

So lets look at these one at a time.

What is Pixel binning?

Well essentially when you take a pixel from your image and clump the data from a few adjacent pixels together to make one pixel. Essentially reducing the pixels you have.

Its not quite like simple down sampling because you also use "noise" measurements (how do you measure the signal to noise ratio in a pixel? I'm sure thats a topic all by itself) to make the resulting pixel 'cleaner'.

So for instance in this figure I've tried to represent how each of the grids of 3x3 are 'binned' together to form a "super pixel" from the results of each 3x3 array.

This site has a good definition summary.

Pixel binning refers to the combination of the information of adjacent detectors in a CCD camera sensor to create one single pixel in the recorded image. For instance, a 2 × 2 binning gathers the electrons from a square of four detectors to record them in just one of the image pixels. Thus, the intensity per pixel increases in a factor of (about) four.

So binning down the 41 MP would give us 3864 x 2684; down to about 10MP (or a quarter as you'd expect from a 2x2 array).

10MP is now suddenly not so "jaw dropping" to the average joe.

Now, lets go back to that picture in the sample above and examine it more carefully and reasonably. For starters the image there is only 5MP (2592 x 1944 Pixels [5.04 MPixels]) which means that its been binned down more than just 2 pixels from its native, its more like 3.

Given that one would think that the noise would be quite remarkable, but as you can see in this image below the grain levels there aren't so clean.

we see here from that segment an amount of noise in the areas like hair and skin tones which would take some care in doing advanced image processing noise removal.

So even though we're down sampling and averaging the noise across a grid of 3 x 3 pixels, the pixel binning isn't really making up for the noise.

This (in my view) should hardly come as a surprise because despite all the hollering the sensor is really still quite small.

And its not restricted to this either. Looking at the other section of the image I high lighted we can see some other interesting effects. Actually because BLOGGER shrinks these images to fit the display I suggest you click on it to see it at the "natural" level I'm examining.

This is a 100% sample of the the red channel from that image. Notice the grid forming up around the ear there? Clearly we're seeing artifacting here from the image processing. Certainly its very well controlled, especially in comparison to what you would get with most other phones!

But it simply can't be compared to the results you'd get from a real camera with a bigger sensor such as one of the 4/3 cameras like the Panasonic G series or the Oly Pen cameras.

Makes me wonder what the native 41 MP images would look like. Personally I'd love to see a native 41MP image so I can examine that. I'm sure that would be very informative.

All of this consistent with the sorts of noise that you'd expect to get from a tiny sensor. In fact its about the same as I got on my compact camera back in 2001 ... which was 2/3" (8.8mm x 6.6mm VS this new sensor at 10.6mm x 8mm).

The sensor is apparently a 1/ 1.2" sensor, which places like DPReview now seem to think as being "oversize". Well its sort of tragic that not long ago we all expected cameras to have 2/3 sized sensors and complained when they went down. Essentially you can only call this oversize if you've been ignoring size for the last 10 years and only looking at pixel dimensions ...

So the question which I think needs asking is this: would this camera produce any better image if it was a native 10MP on the same larger chip?

I suspect that the answer is: maybe not

Now, lets have a look at the lossless zoom.

As a concept it sounds attractive, but again there is something less than expected on this. We all know what optical zoom looks like; and we all (mostly) know that digital zoom looks dreadful. But its not only because its loosing pixels that it looks bad. Its because as you enlarge and crop from the sensor (even if you have a magic sensor) you do not get the same look as if you use optical zoom. This is because you do not get the same increasingly shallow depth of field.

Which means that when you use the 808's digital zoom you will also not get the same look as you will with an optical zoom.

But what about magic sensor 'losslessness"

Well that's interesting too as I doubt you are likely to get "lossless" either. As we see above the result of the scale down of that segment from its native 41MP to only 5MP still gave some noise, a less than ideal image. There was noise present still, as well as significant artifacting present. This despite the use of 1:3 pixel binning.

The more you digital zoom, the less binning that can happen and the more noise you'll get. So when zooming you'll only get to keep a crop of the unbinned 41MP, which is why I'd love to see a native image to see what that looks like ... then at 100% we'll see the sort of thing you get.

bottom line

to me is that this camera makes use of a physical approach to noise reduction. The designers hoped that by capturing more pixels to use image processing techniques over it would result in a better image. I am sure that it gives much better 5MP images than does a smaller sensor camera. Perhaps this technique allowed for faster signal processing (an important factor) than applying algorithms over a 5MP capture. Speed is important.

What I'm not sure of is will this technique give better images than a 5MP sensor of the same physical dimensions as this camera?

As always information is missing from these the makers of the gear so these analyses are always done without the benefit of lots of information. Probably Nokia knows the answers to some of these questions, but they're not sharing.

I'll leave you with a final bit of comparison. This image was taken with my small sensor Nokia E72 camera. Both the 808 image sample and this are 5MP.

Here is a 100% crop from that image ...

Its cleaner in noise than that picture above of the woman. While its taken at a bit higher ISO than the image from the 808 was, it certainly doesn't have any of the extra cunning science or engineering in making it better either.

So why couldn't 5MP on a bigger sensor be simply cleaner at higher ISO because its from a bigger sensor? Sure you wouldn't get the zoom, but it remains to be seen how 'clean' that looks.

Be keen to see that...