K5 raw is 14 bit

Started Sep 20, 2010 | Discussions thread
GordonBGood Veteran Member • Posts: 6,312
Re: K5 raw is 14 bit

MightyMike wrote:

so you're suggesting no notable advantage for going to 14bit at this time, but of course you're still examining the file, how does it compare to the K-x and K-7 in stuff like the noise floor and DR

Mike, I've now obtained a little more data in that I have the original ISO 1600 PEF that I opened yesterday and now the "Falk Lumo" (falconeyes on this forum) ISO 6400 DNG file, both from Beta pre-production K-5's. I think that we can likely trust what we see from raw and the measurement that I make if it is true that the Sony 16 MP sensor has the digital data acquisition electronics on the sensor, as the noise should then be purely a function of the sensor and not the camera electronics.

A few more details as follows:

1) There are only 10 black photosites columns on the left border of the image in landscape orientation, not 12, although the processing does skip over to the right an additional two columns and processes the remaining photosites with the final image excluding a 10 pixel border on all four sides of the final image.

2) Raw images from ISO 1600 and lower are offset to a 14-bit level of about 512 and clip to a maximum of 16382 or 16383, but the extended ISO range of 80 likely has a reduced "clipping" level.

3) Raw images for ISO's higher than 1600 are likely produced by digital multiplication as the ISO 6400 file showed a clear indication that the majority of the readings are displaced from each other to only occupy every fourth code value. In conjunction with this, high ISO raw image files for higher than ISO 1600 appear to already have their black offset point adjusted to zero rather than 512.

4) It's hard to tell what the individual pixel black read noise is from just one sample, even though that sample image has some flat black patches, but it looks like the black noise level should be about two thirds of that of the K-x and thus that there should be about another about two thirds of a stop of Dynamic Range (DR) above that camera on a per pixel basis and more than that on a same size viewing basis .

5) I confirmed that the electron well capacity per sensel is roughly the same as that of the K-x, which is very good considering that there are about 35% more sensels in this sensor.

Thus, my overall conclusion is that the K-5 should have even better image quality at all ISO's than the image quality of the K-x in spite of the greater number of MPixels and thus better image quality at the same viewing size. When compared to the "noise floor" and thus DR of the K-7, there is no comparison in that this appears to somewhat surpass the K-x, which already easily passed the K-7.

As to your question on whether there is any benefit of the 14 bit raw file sze, from my preliminary measurements I would say likely none whatsoever, and if the black read noise actually turns out to be a little lower than what I measured, likely only for ISO's of 80 and 100 for the deep dark shadow detail below about ten or eleven stops below the white clipping level - a very dark level that few are concerned with unless one is doing major boosts in extremely underexposed areas/images. Hardly seems worth increasing the raw file size by about 20%.

For those "in the know" who want to decode the new 14-bit raw compressed data, the 14 bit Huffman table just changes to:

{ 0, 2, 3, 1, 1, 1, 1, 1, 1, 1, 1, 2, 0, 0, 0, 0,
3, 4, 2, 5, 1, 6, 0, 7, 8, 9, 10, 11, 12, 13, 14 }

from the 12-bit Huffman table of:

{ 0, 2, 3, 1, 1, 1, 1, 1, 1, 2, 0, 0, 0, 0, 0, 0,
3, 4, 2, 5, 1, 6, 0, 7, 8, 9, 10, 11, 12 }

where the first row is how many codes have a prefix of 1, 2, 3, ... 16 and the second row is the order of the differentials in readings, with two having prefixes of 2 bits, the next three with prefixes of 3 bits, etc. Thus, a channel can change by up to plus or minus 3 with a total code length of 5 bits, by plus or minus 7 with a total code length of 6 bits, by plus or minus 1 with a total code length of 5, by plus or minus 15 with a total code length of 8 bits, and by plus or minus 1 with a total code length of 4 bits, etc., up to plus or minus 8191 with a total code length of 25 bits. For mostly flat textured images, the file size doesn't really get too big until the noise gets to about 8 bits, which happens for most of an image for ISO's of about 2000 and up, where the majority of the codes will take at least two bytes per photosite reading and where the cameras will likely automatically switch back to uncompressed raw image file output. Note that when I say "plus or minus 3" it actually means plus 3 and minus 4 due to the way signed integer arithmetic works.

The Pentax 645D with its 14-raw files are likely identical to the format for the K-5 other than the pixel dimensions. As it's sensor is a CCD, it is even more unlikely that it actually can use 14 bits of raw output with the generally higher levels of black read noise, but I haven't seen a raw image from the 645D to analyse.

Regards, GordonBGood

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