Dynamic Range (Bill Claff): d3s-d4-d800e-5d3

Started Mar 15, 2012 | Discussions thread
ejmartin Veteran Member • Posts: 6,274
Re: An explanation may be emerging

Marianne Oelund wrote:

Marianne Oelund wrote:

For the D3s above ISO 12800, we're getting into the range where ISO increases are achieved by numerical scaling rather than analog gain increases. But there is something else happening from ISO 12800 to ISO 16000 (H0.3), where somehow the noise level hardly budges; I don't presently have an explanation.

I should have looked at this looooonnnng ago, but it wasn't until just this weekend, that I incorporated a DFT feature into my image-analysis utility.

Nonlinear Filtering

Traditional linear filters are frequency-dependent, and we categorize them loosely into low-pass, high-pass, all-pass, band-pass and band-stop (notch) varieties. The low-pass variety (such as the so-called median filter) can be applied for reduction of high-frequency content, including noise, but when this is done, it leaves a tell-tale "signature" on the frequency response curve.

As I've just discovered, though, nonlinear filters can operate without disturbing the shape of the frequency response curve. A good example is Nikon's hot-pixel suppression (HPS) algorithm, which operates by "suppressing" outlying pixel values. Spectral data from the pre-HPS and post-HPS images looks essentially the same, except for a constant reduction of the component amplitudes across the entire frequency range.

Variations on HPS

Typically, the operation of HPS reduces total noise amplitude (signal standard deviation) by about SQRT(2). Smaller effects, such as seen in the D3s above ISO 12800, could readily be achieved by the same algorithm, with just a single parameter adjustment.

Why would Nikon use this? One of the most visibly detracting types of noise that appears in high-ISO images is what some call "salt and pepper" noise - outlying pixel values that are far above or below the range of most of their neighboring pixels. A "tweaked" HPS would handle these fairly well, without disturbing most of the pixel values. Since the HPS algorithm is already coded, applying it with an altered parameter would be simple.

Now, I'll need to roll up my sleeves and get my hands dirty with some pixel values, to see what's going on in detail.

More later.

If there is NR affecting enough pixels to significantly change the observed std dev, it should show up in the autocorrelation of the raw data.

-- 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