Pixel Density is GENIUS!

Started Jul 13, 2008 | Discussions thread
ejmartin Veteran Member • Posts: 6,274
Re: No really, it isn't

RRJackson wrote:

simpy wrote:

Is it that much noisier if you downsample both to the same
resolution? It may be, of course, but I haven't seen any evidence to
think it is. If it is, it goes on the pile of anecdotal evidence
indeed, but please forgive me in pointing out that you have also
conveniently ignored the image samples by John Sheehy that would go
on the pile of counterevidence.

I haven't seen any test charts from Mr. Sheehy. I've seen an
underexposed low-contrast snapshot that he claims came from a sensor
like the one in the Fuji, but refused to expound on the source of the
sensor. I've also seen some underexposed crops from two captures that
he claims as evidence that smaller photosites are superior to larger
photosites in low light. Neither of these presentations has been
particularly well documented or controlled.

Meanwhile this site is a massive repository of camera test data that
anecdotally points to pixel density as a factor.

I have a little exercise for you:

1. Open a blank canvas in Photoshop. Fill with a uniform middle gray tone. (Image> Adjustments> Levels and set the output level to 128)

2. Add some gaussian noise (Filters> Noise> Add Noise... and choose some amount, say 10%). Look at the histogram and note how its width has broadened from a sharp spike at 128 into a bell curve by the addition of the noise. The width of this bell curve measures the amount of noise.

3. Open the Gaussian blur filter (Filters> Blur> Gaussian Blur...). Watch what happens to the histogram as you very the radius of the blur. You will see the width of the noise histogram decrease in proportion to the radius of the blur. Since the width is a quantitative measure of the amount of noise, the noise decreases as the blur is increased.

Lest you get confused about what you should infer from this exercise, let me spell it out to you. The blur in step 3 decreases resolution by averaging over pixels. Noise decreases in proportion to the amount of resolution decrease. If one downsamples an image properly, one decreases the resolution, and noise decreases in proportion to the linear change in image size.

There is little difference in this regard between shrinking the size of the image keeping the pixel size fixed -- which reduces the resolution by the percentage of shrinkage -- versus keeping the image size fixed and applying a blur, which again reduces the resolution by the ratio of the blur diameter to the pixel spacing. Either way one is averaging over pixels of the original image, and that reduces noise.

The conclusion is that noise is resolution dependent. If you compare two cameras having the same sensor size and technologies but different resolutions (pixel densities), the one with the lower resolution (lower pixel density) will have less noise at the pixel level, as the above exercise has demonstrated.

This site always tests cameras at 100% pixel resolution, and so noise comparisons between cameras of different pixel density, even with the same sensor size, are skewed in favor of the one with lower pixel density, because of the result of the above exercise. Properly resample the higher pixel density to the pixel dimensions of the lower pixel density one, and the comparison becomes fair -- and moreover is a truer indication of what you will see when viewing prints from the cameras at the same distance.

For a quick rule of thumb for comparing cameras of different pixel density fairly, divide the noise measurements of each (as they are done on this site) by the square root of the MP count, which is the linear resolution relative to image size. Note that the noise figure of the higher MP camera will be divided by a larger number, and so come into line with the lower MP camera as a result.

When one does this exercise, it becomes apparent that the main factor in image noise is sensor size. The result is largely independent of MP count for a fixed sensor size. Pixel density, which is sensor area divided by MP count, is poorly correlated to noise because both MP count and sensor size will vary from camera to camera, but only one of those factors is tied to noise level.

-- hide signature --
Post (hide subjects) Posted by
BJN
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow