Rethinking pixel size and high ISO noise

[vandergus]

Alright, this should be fun.

Here's my new discovery.

Increasing the pixel density (decreasing pixel size) does not increase the overall noise in an image.

This is a bit of a paradigm shift for me, and I suspect others, so let me explain. We are constantly told by camera review websites that the megapixel race is detrimental to certain aspects of image quality such as noise and dynamic range because smaller pixels are less efficient light gatherers. This is true, but only if you look at the per pixel level. OK, this part is important so I'll say it again.

This is only true if you're looking at individual pixels.

Of course, this is how most sensors are evaluated on the internet. Images are taken at several ISO's and examined at 100% magnification (per pixel viewing). In this situation, a 10 MP camera appears more noisy than a 6 MP camera (all other factors held equal). If you compare noise on a per area basis, however, which is a more correct comparison when the ultimate goal is a print, then there is no increase in noise. So if you viewed the 10 MP and 6 MP images at the same spacial magnification, say an 8" X 12" print (even identical sizes on a computer monitor), the noise would be very similar.

Alright, some data to backup my wild claims. The DXOmark.com website has lots of data comparing sensors from different cameras. Let's look at some examples. I tried to pick cameras that were from similar generations to minimize the difference in the image processing electronics between cameras. First, the Nikon D3 vs. the Nikon D3x. The D3 is often sited as the ultimate in high ISO performance while the D3x has superior resolution at the expense of increased noise. When looking at the pixel level, this statement holds true, but DXOmark also has the option to compare how the cameras would look when normalized the same print size.

per pixel comparison


per area comparison


They actaully have remarkably similar performance when compared on a per area basis. I'm still working through the rationalizations in my head so I don't really want to try to explain exactly why this is. This is where you guys help me out.

Another comparison, the Nikon D3x and the Nikon D90. The two sensors have roughly the same pixel density (i.e. pixel size) but the D90 is an APS-C sensor (half the area for the D3x sensor). Lets see what the DXOmark data says.

per pixel comparison


per area comparison


As expected, the SNR on the pixel level is very similar but the D3x wins when compared on a per area basis. In this case, I think it's because the image from the D90 sensor has to be enlarged more than the image from the D3x, but like I said, still working my brain around this.

So what does it all mean? Well I think the mantra of "More pixels means more noise" must be reconsidered (there may be other reasons to limit pixel count but let's try to keep this conversation about noise). I still think larger sensors have less noise, but now think that it is due more to the sensor size than the pixel size.

Thoughts?

P.S. This whole thing is really just to indulge my geeky side and will probably be of little help in actually improving anyone's photography.

[kirbinster]

I get lost in all the technical stuff But the real world 10mp D40x had more noise in real world pictures than the realworld 6mp D40.

[OsmosisStudios]

And the real-world D3x has alot more noise than the D3, even though they're both VERY good in that respect.

There was some talk, specifically, of the D3x's pixel density. It has the same density as the D300, D200, D90 and D2 series, just over a larger sensor. While yes, this meant that it had a higher noise levels at 100% of it's 24.5mp, when the image was resized to the lower 12.3mp, it was almost noise free, and I think that's what Fletch is kinda trying to get at.

With my D80, if I take a shot at iso400 and re-size it from it's native 10mp to the D40's 6mp, then the noise that is in the original image is "lost" in the re-sizing algorithm because its just too small to be counted.

[jdepould]

Comparing a downsampled image to one straight off the card is disingenuous. If you downsampled the D3 image the way you did with the D3x, it would have better apparent noise characteristics.

[fletch]

Originally Posted by OsmosisStudios

and I think that's what Fletch is kinda trying to get at.

Err - No I'm not, I haven't waded into this discussion yet!

Originally Posted by jdepould

Comparing a downsampled image to one straight off the card is disingenuous. If you downsampled the D3 image the way you did with the D3x, it would have better apparent noise characteristics.

Yep but you don't end up with a similar image. What Vandergus is saying is that if you shot the same photo at the same ISO on a D3 and a D3X in the originals the D3 would be a noise free 12mp and the D3X would be a slightly more noisy 24.5mp. If you downsample that D3X image to 12MP then you would end up with the same nosie levels as in the D3 shot, and the same number of pixels.

In practice I'm sure it depends very much on the method used to downsample.

They discussed this at great lenght in a DP review blog,

http://blog.dpreview.com/editorial/2...mpling-to.html

[FoG]

Ok you lost me at pixel ... No seriously I get where you are coming from but I reckon they would have covered that mase personally I think from looking through images 10-12 mega p[ixels is enough in the hands of someone who knows what they are about anymore and something has to give .

[vandergus]

Kirbinster: Ultimately, using personal experience is the best way to make a determination about something like this. But I have to ask, were you comparing images using %100 percent crops (viewing at the pixel level) or using images of equal size. If %100 crops were used then the D40x is being put at a disadvantage because its image is being enlarged more than the D40 image.

Everone else: Downsampling is one way to look at it but I want to avoid conversations about downsampling algorithms and the like because it's not necessary to illustrate my point.

What I'm getting at is that comparing the noise performance of sensors of like format (sensor size) but different resolution shouldn't be done from %100 percent crops. Doing this with cameras like the D3 and D3x is like trying to compare the sharpness of two images by making a 12" X 16" print from one and a 16" X 24" print from the other. Well, the image used to make the 16" X 24" print appears less sharp because it was enlarged more. This is obviously an invalid comparison.

Noise behaves in a similar way. It scales with image size. I you have a noisy image it appears less noisy when viewed at a smaller size just as an unsharp image will appear shaper when viewed at a smaller size.

[Gobae]

If I'm understanding the argument correctly, it's like looking at population statistics "per capita" instead of the total number. The only reason there are more noisy pixels on a D3X is because it has more pixels (overall) than the D3.

Interesting. I'm surprised there isn't a "per pixela" rating.

[fletch]

Originally Posted by vandergus

Noise behaves in a similar way. It scales with image size. I you have a noisy image it appears less noisy when viewed at a smaller size just as an unsharp image will appear shaper when viewed at a smaller size.

Yep, I don't think there is any doubt that downsampling reduces noise. The question is does it reduce it by the same amount as simpling using a sensor of the same size but less pixel density.

The DP reveiw article seemed to sugest that it didn't but wasn't definative. The only way to test this would be to get two, very similar, cameras with different pixel densities. A D3 and a D3X would be perfect but as nobody is that rich a D40 and a D40X should do the trick.

And since this is just about noise I definately won't mention anything about dynamic range or continuos shooting speed, file size, resolution, ISO range etc!

[vandergus]

The only way to test this would be to get two, very similar, cameras with different pixel densities. A D3 and a D3X would be perfect but as nobody is that rich a D40 and a D40X should do the trick.

I hate to do this, but the best example I could find of a comparison of actual pictures (instead of sensor data) was at Ken Rockwell's site. Check out this comparison of the D3 and D3x at ISO 6400. What's funny is that he actually gets the crops right. He uses a %100 crop from the D3x and a %142 crop from the D3, which would result in equivalent image sizes. The character or pattern of the noise is definitely different between the two. The D3x has a much finer grain, as you would expect. But I don't think I could say that the level of noise is much different between the two. You can draw your own conclusions but I think this supports the idea presented here.