If I understand it correctly then, especially the bit about RAW
conversion, wouldn't this type of approach be the most effective if
done by the same people who wrote the "de-mosaic" algorithm, since
they would know exactly what kind of weighting is given to each
colour? Seems to me like the best spot to do this would be right in
C1 itself. I wonder if the built-in sharpening in C1 takes anything
like this into account? Methinks MIchael Tapes and Co. should read
this article and see what they think...
I agree with your point, Neil. One issue with Chaney's method is
that many demosaicing algorithms already include some compensation
for the relative densities of 3 color filters in the Bayer pattern.
Without knowing how an image was demosaiced, it's easy to under- or
overcompensate. At least this feature is classified by Mike as
experimental and includes a slider to control the magnitude.
Another issue I have with the article is the continued promotion of
the idea that color resolution of a Bayer sensor is somehow 1/2 of
what the pixel pitch would predict in each direction. This is the
same argument used to bolster the Foveon stacked-sensor scheme, but
isn't entirely accurate for two reasons.
Most real-world images don't exhibit high-frequency variations in
color. They may look like they do, but, if you convert them to LAB
mode and look at the luminance vs. the A and B channels, you'll see
that most of the "detail" is luminance detail and that the color
itself varies much more slowly. A color growing darker and lighter
rapidly can look like its changing rapidly, but its really just a
luminance variation.
Because the R, G and B filters have considerable spectral overlap,
each pixel generally records something even if the color doesn't
match its filter type. A red pixel will show some response to a
color that might look green to the eye, for example. This means
that luminance information can be found in every pixel. If you read
articles on demosaicing, what this amounts to is that the algorithm
has to pick apart the luminance and chrominance information, which
interfere with each other. An algorithm does this not by just
averaging together adjacent pixels of the same color, but by
involving all surrounding pixels in the calcuation of the missing
color channels at any given pixel location. It's not a collapse of
four physical pixels into one output pixel, but rather the use of a
set of data to create another larger set of data that is, in some
sense, a best fit to the subset within the constraints of how a
natural image tends to behave. Good algorithms can achieve
luminance resolution, though, which amounts to perhaps 80% of what
you'd expect from pixel pitch.
Because of this, resolution charts that intentionally consist of
alternating stripes of different color misrepresent the way color
varies in the real world and exaggerate the problem. One can think
of a Bayer sensor as being a form of data compression, taking
advantage of the slowly-varying behavior of color across an image.
6Mpixels worth of raw sensor data ends up yielding a very high
quality full-color image in the end. A 3Mpixel Foveon sensor may
match the quality (somewhat less luminance resolution, somewhat
better color resolution), but requires 50% more data to be captured
and routed off the sensor chip.
As such, its success depends on whether or not its recording the
type of images that its expecting. If demosacing algorithms are
designed around natural images, they won't work very well with
artificial charts. An analogy is JPEG. This compression scheme can
work quite transparently on natural images but produces far worse
results on, say, line art. Other compression schemes, such as RLE,
would work much better there. The success of a scheme depends on
optimizing for the intended data.
I think the biggest issue with the Bayer sensor is that it requires
a relatively aggressive antialiasing filter to prevent Moire
patterns, which demosaicing doesn't handle well and reproduces as
false color patterns. Once you put such a filter in place, though,
you are also blurring the green channel by the same amount, so I
don't know why Chaney's equalization scheme should even work.
David
