Digital Noise vs. Film Grain (Part 1)...

I know that many of you are bored with these discussions on noise, but I think it is important to define an acceptance level, since in another thread it was made quite clear that this is one area that digital has generally surpassed film as it is roughly equivalent to grain. It is also one of the main distinquishing parameters between images from consumer level digicams and DSLR's.

First, it is important to define a specification for the level of noise we can barely detect, and one the we find marginally acceptible. It is quite easy to use a image editing program such as Adobe Photoshop to add Gaussian noise to different output levels to determine the level at which we can barely detect such noise when fully adapted to a full dynamic range from darks to whites over a full image (no microscopes/huge zooms). We find that we are most sensitive to noise in the medium grey area (in a linear luminance sense), which for a standard image on a standard gamma monitor is about 64 out of 255. We find that we can just detect a Gaussian noise level of a standard deviation of about 1.5 bits, which is also theoretically correct, as the extent of the Bell curve ranges up to about plus and minus 3 levels to the 99th percentile level, and Colour Science teaches us that the Human Visual System (HVS) is only sensitive to about 2.5 levels of change out of 256 at this level. In the same way, we can establish that about double this level, or a standard deviation of 3 bits, is somewhat acceptable, especially for making prints where the limited dynamic range and number of colours limits our ability to perceive as fine a distinction in level change. These limits are for true Gaussian noise, and not for noise reduced noise, which usually has a larger "grain" size. Using a median filter with a radius of 1 pixel to simulate this, we find that the equivalent limits for this noise reduction pattern is about half to one third of those established before. This is likely for two reasons: the HVS automatically seems to apply its own noise reduction averaging for a normal Gaussian noise curve, and the averaging type of noise reduction limits the maximum excursion of the noise but clumps more of the noise into larger areas that the HVS can still perceive. These techniques also reduce apparent detail, and even when sharpened to give the perception that detail is still there, to leave a plastic watercolour type of effect in the images. Thus, it is only noise reduction techniques such as wavelet analysis or adaptive frequency domain filtering as used by NoiseNinja or NeatImage that can truly reduce noise while retaining detail, but these are too slow to be implimented in digicams through firmware, and hardware implementations haven't appeared yet.

Next, we establish a grain/noise level for scanned film as measured from the many sets of comparative studies posted on the Internet. We find that the level of grain for what is considered to be an excellent film in Fuji Provia 100f is about 4.5 to 5 bits of standard deviation in a scanned image, with a fairly coarse pattern (at least by digital camera standards). We find that this is somewhat bigger that the acceptable limit as established above, although perhaps some of the perceived grain comes about due to the scanning process, especially image sharpening. It indicates that film grain is equivalent to most digital cameras used at their top ISO sensitivity setting or one less than this, and that I may be too fussy with my chosen limits above.

To be cont'd...