2 Noise Distribution Sample The noise in a digital image sensor comes from several different sources.  The largest component in modern sensors is due to the Poisson statistics of incident photons, proportional to square-root of the intensity.   In very low light levels other noise sources become apparent.  This is an examination of noise data obtained from a Canon EOS 20Da. Here is a portion of a frame from an exposure taken through a telescope.  The exposure was 1-second at an effective aperture of f/9.9 (Takahashi CN-212, at 2630mm with focal reducer).  This is the shortest exposure that utilizes Canon's noise reduction method of obtaining a successive dark frame and processing.  The ISO setting was 800, ambient temperature around 0 C0.  A 16-bit linear tiff file was obtained from the camera raw file via the Canon EOS Viewer Utility, and a representative 560x440 pixel section obtained for analyzing. Albireo.noiseSample.1-3339.tif A section from the background of a 1-sec exposure at f/9.9, ISO-800.  The values have been multiplied by 255. Only even-valued levels were observed in the data, so a histogram was created where values were split evenly between adjacent bin pairs.  Here is the low end of the histogram along with a gaussian model that matches mean and variance. This may look like a reasonable match (apart from the obvious numeric scatter near zero), but I am mostly interested in the noise levels in the tail, which may have a strong impact when the image is scaled as part of a high dynamic range composite.  To see the behavior in more detail, a log is taken.  It is obvious that the normal distribution falls off rapidly compared with the observed noise. We obtain the cumulative distribution functions by summing (which also helps show the behavior of the low amplitude noise). Subtract from unity and take the log to examine the tail of the distribution, clearly showing the difference between the image data, and the gaussian model. Next, try to fit a composite noise model to these statistics. (Next page.)