500d density

Started Apr 2, 2009 | Discussions thread
Veteran MemberPosts: 6,274
Re: Noise scales
In reply to Manehead, Apr 10, 2009

Manehead wrote:

Daniel Browning wrote:

[ snip ]

I'm not sure if the post you linked did "thoroughly debunk" my
downsampling claims.

Yes, it did.

As far as I can tell, that post mainly explains why the downsampling
didn't remove as much noise in the medium and large grain images
compared to the fine grain. It basically reiterates the same reasons
for this as those given by the original blog post, i.e. because the
noise grains are larger than 1 pixel, and appear in the form of
"blotches" of several pixels in size.

The downsampling didn't remove any noise because the "coarser grained" samples were generated by upsampling the "fine grain" sample; I showed this in a followup post:


There is no noise to be removed by downsampling in this example because upsampling and then downsampling back to the same size won't have any effect (if done properly). The upsampled image has the same noise on the same scale in the upsampled image, and no noise on scales finer than the pixel pitch of the original image. There is no fine scale noise for the downsampling to remove, because the images were generated from a source that had no anything on those scales -- no noise, and no image detail either!

But this is orthogonal to what I think is the central issue, that noise at equivalent image scales is the same. This point can be made independent of any resampling, by direct examination of the two images at comparable spatial scales relative to the frame size. People often use the resampling argument to discuss the relative amounts of noise in images, but (as Phil's blog post illustrates) it is a red herring. Resampling introduces a whole host of additional variables that have little to do with the image content of the two images one is wanting to compare. Instead one can simply examine the noise spectra of the two images without resampling, which is what I did in some of the posts Daniel linked to (or similar posts in the same threads).

This "blotching" effect is the critical element that determines the
accuracy of my my proposal about downsizing not reducing as much
noise as expected. Based on both the original blog and the linked
post above, if the size of the noise are generally larger than 1
pixel, then the noise reduced will not be proportional to the amount
the image is resized. However, if it turns out that there is no
"blotching" and the noise levels in each photo cell is completely
independent of is adjacent cells, then downsizing an image will
effectively reduce the image noise in a linear fashion.

The problem is that this way of viewing things is fixated on noise at the pixel level, which is comparing apples and oranges (image properties at disparate spatial scales, rather than the same spatial scale). The blog post example illustrates this perfectly -- the noise at the pixel scale of the upsampled image is depleted, because there was no noise in the source image at scales finer than its pixel level. Instead, it is more appropriate to compare the pixel-scale noise of the source image, to the noise in the "coarse grain" image at the scale (larger than pixel-scale) that corresponds to the pixel scale in the source image. Those will be the same with good resampling technique.

As far as I saw, the only section of the linked post that attempt to
address the reasons behind this "blotching" is near the end, and I

"The size of the "grain" above is mostly an indication of how much
noise reduction has been performed on the image during conversion and
post-processing; having examined the power spectra from just a decent
demosaic process, demosaicing has the effect of lowering the slope of
the power spectrum in the upper half of frequencies, but the power
doesn't tail off nearly so dramatically as in the above samples."

This paragraph implies that blotching may occur as a side-effect of
jpeg conversion/compression (i.e. "post-processing"), and therefore
resizing a JPEG image will not yield the theoretical amount of
reduction in noise (this reinforces my final point in the above
post). However, it does not address the question of whether RAW files
also exhibit blotching.

The RAW data does not, in any current production camera I am aware of (including the A700 with the latest firmware and NR off), have a depletion of noise power at the finest spatial scales (which would show up as "blotching").

If RAW files really are the unprocessed "power spectra from just a
decent demosaic process" then they should not show any blotching,
unless the blotching is inherently caused by the hardware limitations
of the sensor. However, there has been doubts raised as to whether
the RAW files generated from current DSLRs are really "raw". For
example, the RAW files from the Sony A700 were originally created
with noise reduction already applied, which was later rectified by a
firmware update. This goes to show that the original sensor data does
go through some kind of processing before being output as RAW. It may
be possible that this processing is the cause of these blotches
refered to in the blog post, or they might have simply been refering
to JPEG images rather than RAW.

See above. Canon RAW files are not "blotched". (Some might argue that Canon's recent hardware releases have been "botched", but that's a different story

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