Burst + High ISO to improve noise

[Guillermo Luijk]

This is an exercise I wanted to do time ago, but never found time and energies. This afternoon I did it, and although the result was not so good as I expected I wanted to start a discusion.

I wanted to compare the noise in a regular shot at base ISO100 vs an statistical averaging of several shots at higher ISO values. If the total exposure time was kept the same, the concept could be applied in a digital camera with a fast burst operation: the user clicks once, but the camera shoots several times in a quick sequence that will be averaged to improve noise. ISO is raised as needed to keep the total exposure time into the required shutter speed set by the user for the particular situation.

Please don't compare or think of applying this to present cameras, this is just a concept idea for fast burst cameras that could come in the future.

I shot the following scene with a Canon 350D on a tripod:



- First I shot it at ISO100, 2s
- Second I shot it 8 times at ISO800, 1/4s

The total exposure time was the same, 2s. Then I compared noise of the first shot (ISO100) vs two different statistical averaging of the other 8 shots: simple Average and Median. Surprisingly median was less effective in reducing individual impulsive hot/crazy pixels than simple average; that makes me think these hot pixels remain at the same locations in most of the 8 high ISO shots (I'll try a simple hot pixel elimination routine tomorrow).

There is a reduction in noise in the averaged images, but not spectacular at all. Anyway I think the concept can be interesting.







Images are B&W because I wanted to average undemosaiced data. I arbitrarily picked up the B channel of the RAW files and did the averaging without even clipping the black level in case Canon's non zero black level could mean an advantage in the end for this kind of noise reduction.

_________________

Application for HDR: the same idea of keeping more than one shot into one affordable total lapse of time, could be applied for in-camera HDR shooting.

Think of a landscape that has to be shot at 1/24 or faster because we are using a 24mm lens. The user sets shutter 1/24 but the camera could do two shots a couple of EV's apart: one at 1/30 plus one at 1/120. Total exposure time would be 1/30+1/120=1/24, enough to freeze our handheld shot at 24mm. The 1/30 shot would provide low noise in the shadows and 1/120 would provide the highlights. The blend would be automatically done by the camera without the user noticing. The 1/30 shot would only display a slightly higher noise in the shadows than a single shot at 1/24, nothing compared to the advantage of preserving 2 extra stops in the highlighs.

Regards.

[ejmartin]

According to Roger Clark's data tables, the read noise at ISO 800 is a little over 4x smaller than the read noise at ISO 100 on this camera.  On the other hand, by stacking 8 ISO 800 images, you get 2Sqrt[2]~2.8x the read noise of ISO 800.  So one should only expect in your test a little over 1/3 stop decrease in read noise.  Random pattern noise should be somewhat improved.  Shot noise is a wash.  So I would not have expected to see as dramatic a difference as there is between single shots of fixed total exposure (a test you have done before).

The big advantage of stacking is in averaging the random pattern noise, and any benefit that comes from the lower read noise in electrons at higher ISO minus the detriment of increased noise from multiple exposures.  ISO 800 is about optimal in this regard, (for instance a good setting for astro work).

[Ray]

This is an exercise I wanted to do time ago, but never found time and energies. This afternoon I did it, and although the result was not so good as I expected I wanted to start a discusion.

I wanted to compare the noise in a regular shot at base ISO100 vs an statistical averaging of several shots at higher ISO values. If the total exposure time was kept the same, the concept could be applied in a digital camera with a fast burst operation: the user clicks once, but the camera shoots several times in a quick sequence that will be averaged to improve noise. ISO is raised as needed to keep the total exposure time into the required shutter speed set by the user for the particular situation.

Please don't compare or think of applying this to present cameras, this is just a concept idea for fast burst cameras that could come in the future.

Guillermo,
There is a use right now for this feature of stacking multiple shots using either mean or median rendering modes. Supposing you are not carrying a tripod and want to take a still-life in poor lighting where flash is not appropriate. Lets suppose the minimum shutter speed to prevent camera shake spoiling the shot is 1/50th at ISO 3200, at the aperture you've chosen for appropriate DoF.

Cameras like the 40D, 50D and 7D have a reasonably fast frame rate to allow a burst of 6 or 8 frames within a second or less. Provided you can hold the camera reasonably still for a second, the auto-alignnrnent feature in Photoshop can handle any small misalignmrnt perfectly.

From memory, trying it with the 5D which unfortunately has a rather slow frame rate, I was able to get at least a full stop improvement in both noise and resolution. This is useful when merging to HDR is not an option because the shutter speed of the longest exposure is too slow to freeze camera shake.

[EduPerez]

Is this related to the stacks astronomers do? Thanks.

[Ray]

Is this related to the stacks astronomers do? Thanks.

I think it probably is, but I'm no astronomer. In CS3 Extended (and presumably CS4 which I don't have), there's an option of stacking a number of identical shots whereby the best pixel from each frame is selected to produce a lower noise and higher resolution resulting image. It really does work. Of course, no two shots are really identical, and stacking takes advantage of this.