k7 vs k20 - noise analysis

Started Jul 16, 2009 | Discussions thread
ForumParentFirstPreviousNextNext unread
Flat view
Oleg_V
Contributing MemberPosts: 627
Like?
k7 vs k20 - noise analysis
Jul 16, 2009

After the first discussion of the subject http://forums.dpreview.com/forums/readflat.asp?forum=1036&message=32352733 I have performed more extensive study of IR samples and comes to some important conclusions that agree well with Gordon findings.

First a little bad thing is that k7 noise is actually worse than that of k20. The difference is probably 1/3 step on ISO scale. The second observation is pretty much good - unlike k20 I failed to see any sign of NR in RAW files at ISO1600 and below. Thats simply great IMO, I'm starting to love pentax again!

Now I'm going to present the results for those who are interested. I start from ISO3200 data extracted from the following IR samples
http://www.imaging-resource.com/PRODS/K20D...00V1.01.DNG.HTM
http://www.imaging-resource.com/PRODS/K7/K..._NR_OFF.DNG.HTM

The square 400x400 area with left top corner at (20,20) were chosen for analysis as being more or less plain gray. First I've extracted channel data using 'dcraw -d -4'. After that they were processed by handcrafted tool calculating noise autocorrelation function as a function of distance that is measured in bayer cell units (2pix). Below you can see the results for red/blue channels

So the NR trace is clearly visible at least at distance 1. Moreover, the k7 NR is almost 2 times stronger at that point. Some difference in NR strategy is seen in green channel as well (see below)

The inter-cell autocorrelation curve looks the same but what is seen in k7 only is intracell green channel averaging (we have 2 green pixels per cell).

Now lets look at how NR depends on ISO. On k20 the NR trace is detectable down to pretty low ISO (see below). Actually we can't tell at which ISO NR is off completely due to limited accuracy. Probably thats why K20 images always looked soft to me pixelwise ..

Not so with k7. The autocorrelation NR trace drops to noise flour already at ISO1600 (see below).

Now lets consider the noise levels themselves. Since we don't have much data the simplest way is to qualify noise using 'effective gain' parameter.

Provided that the noise is pure photon one can calculate the gain in output code units per photon by simple formula G=N*N/S. The result depends on ISO only and not depends on signal level. In reality the result will be affected by NR (decreasing it) and dark readout noise (increasing it) but for the sake of simplicity such 'effective gain' is pretty useful. Below you can see how it depends on ISO for k7 and k20 (gain is related to 16 bit range).

So you see the k20 have lower noise overall but remember - it is the result of NR applied. Still I think without NR k20 noise would be lower since after more aggressive NR at ISO3200 it sill higher at k7 samples. Though the difference is not that huge. Unlike the k20 the point where NR is started is clearly visible on k7 curve.

So, my conclusion for those who happened to read all this stuff. I think k7 is a good camera delivering slightly more noise but more details as well. Good point from engineers after all. As for ISO3200 and above - I don't care since they are barely usable for detailed images.

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