More E-PL3 noise results...

Last night, I took some more dark frames to try to characterize the noise.  Selecting only the "warm" pixels from the 60s dark frame that I analyzed in the last post, I looked at the signal distribution from those same pixels in a sequence of dark frames with exposures from 30 seconds to 1 second.  The first figure shows that the mean and width of the warm pixel distribution scales with exposure time, and that (based on the 30s exposure) the selected pixel population comprehensively captures the pixels making up the warm pixel population in all exposures.  In other words, the problem pixels are a fixed population.  Furthermore, with this data set, I find that the mean of the warm pixel distribution is linear in time (see 2nd figure).

Warm pixel signal (selected from 60s dark frame) for a sequence of shorter exposure dark frames.  The blue histogram is the full distribution (all pixels) from the 30 s dark.  Note that the selected pixels from the 30 s fill the anomalous noise peak of the full distribution.  This shows that the problem pixels are a fixed population.

The other quick experiment I did was to take three 60 s darks in sequence.  This time, I'm showing the full distributions with the Y-axis log scaled.  Between the first and second dark frames shown here, I took a few shorter dark frames.  The second and third are taken directly sequentially.  The data here shows that the mean of the warm pixel distribution (dark current) is temperature dependent.  This is not a surprising result, except that for the bulk of the pixels, the signal stays put at ~64 ADU, though the main peak does broaden for the later frames.

The upshot of this analysis is that "noise reduction" via subtraction of a dark frame taken immediately after the image is the best way to combat this noise.  The 1-30 second sequence indicates that this becomes important for exposures longer than 8 seconds, though this should be taken as a temperature-dependent statement.

Another interesting note (not shown here, so you'll have to take my word for it).  Within the warm pixel distribution, the noise appears to be uncorrelated.  That is, if I subtract the signals from two frames, the width of the difference's distribution is the quadrature sum of the two distributions.

One more interesting note:  I would expect the variance of the distributions to scale linearly in time, but it does not.  It's faster than linear.  I can't explain right now.

Long exposure noise analysis inspired by the E-M1 noise problem

(Peculiarly, blogspot only shows previous posts, but not subsequent ones.  Here is the follow on to this post: more epl3 noise results)


The noise issue that has come to light recently with the Olympus OM-D E-M1 in long exposures caught my attention last night, so I decided to take a look at the output from my camera, an Olympus E-PL3.  While the E-M1 and E-M5 use a 16 Mpix Sony sensor, my camera has a Panasonic 12 Mpix sensor.  In both cases, the technology is CMOS, not CCD.

So far, I've only taken one image and reduced the data from that single frame:

• 60 sec exposure
• ISO 200
• IS off
• Noise reduction OFF
• Noise filter OFF (not sure if this matters w/ raw image)
• f/22, lens capped to minimize stray light. 

I converted the raw file into a FITS file using dcraw
% dcraw -D -4 -c file.orf | pnmtofits > file.fits
and then imported the file into ipython using the astropy library.

On my sensor, rows 4056 to the end have some overscan garbage, so data in those rows were removed from further analysis.  In the first figure, below, I histogrammed the noise from the sensor.  The main peak is at 64 ADU and does not extend much beyond 100 ADU (top panel).  In the middle panel, the Y scale is expanded, and the second noise peak is evident.  This might be analogous to the noise seen in the E-M1.  The bottom panel shows the noise binned into 50 ADU chunks.  All of the information in the top two panels fits into the 5 left-most bars of this histogram.  There are two more noise peaks, at 500 and 1200 ADU.

I found 93 "high count" pixels, mapped below.  Looking at their spatial distribution and density, I think they are unlikely to cause any imaging problems.

The second noise peak, is more interesting.  It peaks at 179 ADU and is done by ~230 ADU.  There are 37000 pixels in the 130-230 ADU range.  As you can see below, their distribution is (at least by eye) uniform and the worst offenders tend to cluster in the corners.

 

I have no idea if this is the same problem that the E-M1 has, but if anyone wants to send me some E-M1 dark frames, I'd be happy to run the same analysis on them.  I'll probably take a few more frames to verify my hunch that this is a fixed-pattern problem, which is why noise reduction can kill it off easily.