Noise buzzing in the ears

Discussion in 'General Technical Discussion' started by Iliah, Nov 10, 2005.

  1. Iliah

    Iliah

    Jan 29, 2005
    nowhere
    Noise is big deal. Low noise at high ISO is paramaount. Let's take Kodak. We all know that Kodak cameras are unusable at ISO higher then 250 (or 400, or unusable at all above basic ISO 80 - depends on whom you ask).

    Here is a link to 11Mp Kodak SLR/n shot taken at ISO 640, without *any* noise reduction applied, but with pretty extreme sharpening instead:

    http://www.rawmagick.us/CatRaw.jpg
     
  2. Link doesn't work... at this time.

    Woody
     
  3. heiko

    heiko

    May 15, 2005
    Israel
    Hi Iliah,

    I'm not familiar with the Kodak SLR, but I would be very happy to get this result out of my D70. Some sharpening artefacts can be seen at full size, but the noise is only single pixels that don't intrude.

    Can you do the same test with a D70 image (I got lots of them with pretty bad noise @ ISO 800 and above)? I assume you used your Rawmagick on that.

    Very nice result!!
     
  4. Gale

    Gale

    978
    Jan 26, 2005
    Viera Fl
    Thats Fantastic Iliah.
     
  5. Iliah cheats, he knows what he is doing :biggrin: :biggrin: :biggrin:

    This is done via RawMagick, of course, but not applying any of the Noise Reduction bits? To what do you attribute the smoothness of the dark background? Is the noise quite apparent when this image is rendered other than with RawMagick?
     
  6. Iliah

    Iliah

    Jan 29, 2005
    nowhere
    Dear Heiko,

    Please send me your ISO 800 NEF to see what is wrong.

    Dear Gale,

    Thank you, and thanks to Kodak. I hope to see their new sensor in Pentax camera soon, chances are as early as this spring. Kodak chip in Leica is also doing very good IMHO.

    Dear Bill,

    Hm.... The image was from Kodak site, downloadable as a sample :) You can try it in ACR or any other converter for that matter.

    Cross from Dan's Color Theory list:

    "Even a 16 bit INTEGER image only has 16 tones to describe the darkest stop in a 12 stop image, while for the brightest stop, 32,768 tones are available. This is the opposite of human vision which is more sensitive to shadow detail than to highlight detail. A 32 bit INTEGER image provides more tones but has the same limitation of having a disproportionate amount of tones for the highlights. 32 bit FLOATING POINT images address this issue by making more efficient use of the 32 bits. Instead of using 32 bits to describe 4,294,967,296 integer numbers, 23 bits are allocated to a fraction, 8 bits to an exponent, and 1 bit to a sign, as follows:

    V = (-1)^S * 1.F * 2 ^ (E-127), whereby:
    S = Sign, uses 1 bit and can have 2 possible values
    F = Fraction, uses 23 bits and can have 8,388,608 possible values
    E = Exponent, uses 8 bits and can have 256 possible values

    Practically speaking, this allows for an almost infinite number of tones between level "0" and "1", more than 8 million tones between level "1" and "2" and 128 tones between level "65,534" and "65,535", much more in line with our human vision than a 32 bit integer image.

    Because of the infinitesimally small numbers that can be stored, the 32 bit floating point format allows to store a virtually unlimited dynamic range. In other words, 32 bit floating point images can store a virtually unlimited dynamic range in a relatively compact way with more detail in the shadows than in the highlights and take up only twice the size of 16 bits per channel images, saving memory and processing power. A higher accuracy format allows for smoother dynamic and tonal range compression."

    Add good demosaicing algorithm to the mix.
     
  7. Flew

    Flew

    994
    Jan 25, 2005
    Alabama
    Iliah,

    I'm confused (not a hard thing to do :rolleyes:). Is this result based on:

    A) A great new low noise sensor from Kodak?

    B) A new way of processing RAW files in RawMagick?

    C) A combination of A and B.

    Whatever the the case, this is amazing. If the answer is B, then I need to get my copy of RawMagick upgraded and give it a try. If I could get results this good from my D2H at even ISO 500 or 640, I'd be in heaven. :biggrin:
     
  8. Iliah

    Iliah

    Jan 29, 2005
    nowhere
    Hello Frank,

    Please wait a day or two - current release is unstable.
     
  9. Flew

    Flew

    994
    Jan 25, 2005
    Alabama
    Iliah,

    I'll be glad to wait, but as I said, if you can make images from the D2H, at ISO gain settings at or above 500 look like the image that you posted in this thread, I want it. :smile:

    Let me (and all of us) know when we can get the next version. :wink:

    Thanks,
     
  10. TOF guy

    TOF guy

    208
    Mar 11, 2005
    Hi Iliah,

    Well I'm going to play the devil's advocate.

    A 12-bit image using 32 bit integers leaves approx 1000000 steps between the two darkest tones recorded (0 and 1 in the original image). That's should be much more accuracy than we ever need. How accurate is the sensor ? How many levels our eyes would discern between these two dark shades of grey ?

    Sure looking at the big picture - so to speak - of a 12-bit image mapped into 32 bit integers, there is a lot of encoding information wasted in the highlights. So that information doesn't help. It doesn't hurt either. As long as there is enough level gradations in the darkest areas of the picture.

    Calculations using 32-bit integers would use as much memory as the 32-bit real numbers. But it may save computer resources in another area: aren't integer operations much faster than their floating point equivalents ?

    Can't wait for that mysterious release !

    Thierry
     
  11. Iliah

    Iliah

    Jan 29, 2005
    nowhere
    Dear Thierry,

    In some programs raw processing is done through many (adaptive, that is not the type that can be pre-calculated and converged into LUT) iterations, starting from noise reduvtion based on masked frame, through vignetting control, and up to visualisation. All this calculation stress the shadow portions extremely.


    Apple has some very nice hardware, and videochips are getting some nice floating point accelerators.
     
  12. TOF guy

    TOF guy

    208
    Mar 11, 2005
    Hi Iliah,

    Thank you for the explanation. I've learned something new (so I'm glad to have played the roled of the devil's advocate) :biggrin: !


    Everything is standardized (or about to be) on Intel hardware these days. Have you tested floating point (with sse/sse2 instruction sets enabled) and integer performance (I don't know how it would compare on recent hardware myself) ?

    Thierry
     
Loading...
Similar Threads Forum Date
Noise and exposure. General Technical Discussion Aug 16, 2014
Noise adjustment: in camera or not? General Technical Discussion May 27, 2012
A funny thought about noise reduction General Technical Discussion Apr 27, 2012
en-el4a charger making, noise normal? General Technical Discussion Mar 23, 2012