E-M5 has DR 13!

Started Apr 13, 2012 | Discussions thread
Detail Man
Detail Man Forum Pro • Posts: 16,688
Re: More a function of how they define DR

Bart Hickman wrote:

Detail Man wrote:

The pixel-binning techniques that you describe work on uncorrelated spatially-distributed random noise (such as Photon Shot Noise), but to the extent that Read Noise (the dominant component of the noise-floor in specifying Dynamic Range) is not entirely random (but is instead periodic) on the level of (within the area of) the pixels binned, those relationships do not hold to be true ...

I'm talking about post-process pixel binning. If you're talking about some sort of fixed spatial pattern that every image has, I agree that's there, but that won't register on these DR tests (and if they are low frequency enough, most people probably won't care about them.) The Pentax K10d has the pattern noise very bad, but that was unusual. The noise (whatever the cause) even in low ISO images appears to be mostly random to me and does respond to simple low pass filtering (I'm not talking about pixel binning, although that should help as well, it makes aliasing potentially worse.) Taking multiple images and averaging them together also significantly improves DR for the same reason.

I do not doubty that applying spatial-frequency low-pass filtering algorithms (whether they be simple arithmetic averaging by binning, or various down-sampling algorithms) would yield better looking numbers. DxOMarks normalization by the square-root of the total number of samples (pixels) is one example. But it is just a game to be played for the purposes of massaging data ...

Anyway, I was mistaken about how dpreview measures DR--they stop going into the black levels once the noise exceeds the level itself. So it's a combination of the two types of DR I mentioned.

Sounds like the same thing that DxOMark does - setting the lowest measurment level at a point where the noise equals the (albeit small) illumination signal (when the Signal/Noise Ratio = 1.0).

John Sheehy wrote:
There is no camera on the market with uncorrelated read noise.


So maybe I don't know what "read noise" is. I just assumed that was all the noise the signal gets while passing through the electronic channel on its way out to the final image. It's easy to understand that there will be gain and offset errors that vary in a fixed way across the sensor area, although I would've thought stuff like that gets calibrated out. I don't keep track of which type of noise dominates DR in a particular image. All I know is the random noise I see.

It is true that manufacturer's strive towards attempting to make the noise-spectrum of read-noise (when sampled over numerous photo-site locations, anyway) to appear as random as possible so that it approaches random uncorrelated noise that can be reduced by spatial-frequency averaging other low-pass filtering techniques, similar to shot-noise sources.

See: http://forums.dpreview.com/forums/readflat.asp?forum=1018&message=40320601

Pixel binning is only useful to save storage space, or it can cut read noise by some amount if it is done before reading the photosites, in camera (and not with all technologies at all ISOs - binning at ISO 400 would serve no noise purpose whatsoever on an analog CMOS sensor like the ones Canon uses in their DSLRs). IOW, multiple pixels' charges are combined and read out as a single pixel. That would help reduce image-level read noise. However, any noise created by cameras after readout will be worse with hardware pixel binning, because less pixels have the same added noise each.

Again, I'm talking about the "binning" that DR testers do on those grey patches and the claim in the previous post that 12-bit digitizing automatically limit the DR to 12-bits (which it doesn't).

However, on a level (non-massaged) "playing field", such things might be termed as "window dressing" ?

Are you saying binning outside the camera doesn't help? I don't think image processing software bins pixel anyway when downsampling an image--I think it does a regular LPF.

Arithmetic averaging is a crude form of spatial-frequency low-pass filtering. Other down-sampling algorithms provide steeper slopes of roll-off (at the price of inducing "ringing" / "artifact" effects).

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