Too many megapixels? Six is enough for an 8x10.

Started Apr 26, 2009 | Discussions thread
ejmartin Veteran Member • Posts: 6,274
Re: All true essentially

R Butler wrote:

re: ejmartin

I'll try to find a moment to read those posts in more depth.
Ultimately, though, my initial scan through suggests that much of the
disagreement comes down to the per area/per pixel argument. And while
it's by no means flawless, there are arguments in favour of a
per-pixel assessment of image quality and strong arguments against us
changing. Which isn't to say we're incapable of listening to reason.

The problem with the per pixel comparison of cameras with different numbers of pixels in the sensor is that it ignores the fact that noise is a function of spatial frequency. I analyzed the noise power of the 40D and 50D at the time of the blog post, and here is what I found (40D in red, 50D in blue):

As you can see, they are the same at the spatial frequencies that they have in common. The 50D has noise power that extends beyond the Nyquist frequency of the 40D, because it has a higher Nyquist frequency. But for spatial frequencies that both cameras can sample, there is no substantive difference.

To compare the noise of two different cameras one should compare the noise power at the same scale; other measures are methodologically flawed.

What does the pixel std dev of levels in a uniform patch measure? It is the area under the noise power curve. Any camera with comparable noise power, but a finer pixel pitch and therefore higher Nyquist frequency, will have higher pixel level std dev, even though they have the same noise characteristics at comparable spatial scales.

In the 40D/50D example, proper downsampling reproduces the 40D noise power curve (40D in red, 50D in blue, 50D downsampled with PSCS3 bicubic in orange, 50D downsampled with ImageMagick Lanczos in black):

Throwing away the high frequency noise spectrum reproduces the 40D noise power, and measurements of the pixel level std dev shows much the same results.

But downsampling need not have been performed; it introduces a whole host of additional issues into the discussion that are not germane (for instance, that Lanczos resampling is more accurate than bicubic for downsampling). The noise power was the same for comparable scales in the image before resampling; all that resampling did was throw away the high frequency image content, and the high frequency noise along with it, from the finer pixel pitch camera.

Noise reduction (as well as the upsampling Phil did in his example) is a local operation, and therefore only affects high frequency image content. If noise reduction is performed first, and then downsampling, then high frequency noise has been removed before downsampling, and so there is little for downsampling to chop away; you can't remove what isn't there. This still has no bearing on the comparison of noise at comparable scales, which as I have shown above can be done without resampling. Downsampling can be a flawed method of comparison if one doesn't understand the facts about noise power and its frequency dependence.

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