My understanding is that the problem is in the electronics that reads out the sensor. If the problem is more apparent at low ISO, then it is in the electronics (such as the ADC) in the signal processing chain after ISO amplification takes place; if the problem is roughly independent of ISO, then it is the electronics (such as the sensor) located before ISO amplification takes place. Reason being that ISO amplification boosts the signal, so if noise is after amplification the S/N is improved by raising ISO at fixed sensor illumination, while if noise is before amplification then raising ISO boosts signal and noise at the same rate and has little effect.
I\'m no electronics expert, but an educated guess is that the line noise may be radio-frequency noise in the electronics. The pixels are read out a column at a time, at a rate of several tens of MHz (since there are 21 mega-pixels read out in a fraction of a second). Any RF noise would oscillate up and down, whichever direction the noise is on a particular half-cycle, the pixels in that column will all be shifted that way.
Why it appears in some shooting conditions and not others for a given boost of the exposure in a converter is very likely related to the color temperature. Reds and blues are boosted by large factors depending on the color temp of the lighting, so are virtually always \"underexposed\" relative to the optimal sensor operating point, which is typically (but not always) set by the green channel which is closest to the luminance data. You can see how much the R and B are boosted by white balance if you look in the exif data using exiftool or a similar metadata reader. For instance I have a 7D file to hand for which the relevant line in the metadata is (cloudy WB)
Measured RGGB : 518 1024 1024 743
and another for which it is (tungsten WB)
Measured RGGB : 1038 1024 1024 314
In the cloudy case, the R channel has half the exposure (518) as green (1024); and blue is about a half-stop underexposed (743 vs 1024). In the tungsten case, red and green are approximately balanced, but blue is about 5/3 stop underexposed. In each case, to get white balance the R and B channels must be boosted in the converter software by a factor 1024/518, 1024/743, or 1024/314, respectively. So you already start off with a stop or more of \"lifting shadows\" in R and/or B even before you touch the exposure slider.
HTP also intentionally underexposes the image by using 1 stop lower ISO internally in the camera, and the converter internally has to boost the shadows by another stop.
So I would recommend anyone interested in seeing the boundaries of how to expose without generating issues with their 5D2, to look at a sampling of shots to see how much lifting of shadows is needed to show problems for their camera at various ISO, and then look in the exif data to see what the white balance multipliers are. That will give a rough guide to how much amplification of the color channels is taking place. It should be that the problem appears consistently when the combination of exposure compensation and WB multipliers exceeds some threshold. And by examining how light color temp relates to WB amplification of R and B, you will know in a particular shooting situation how much leeway you have in exposure -- how much of the total compensation before pattern noise becomes a problem remains, because some has already been used up by the WB amplification that will be necessary in converting the file.
Nov 15, 2009 at 10:21 AM
Previous versions of ejmartin's message #7778118 « 5D2 banding: what does it take? »
