From: David J Taylor on
"Alfred Molon" <alfred_molon(a)yahoo.com> wrote in message
news:MPG.2643dae57b79e95198c2be(a)news.supernews.com...
[]
> It does, because that is exactly what Bayer does: interpolate missing
> values, by assuming that they are a function of the data in the
> neighbouring pixel. These are guesses, which often are wrong. You cannot
> calculate the value of a random variable.
> --
>
> Alfred Molon

If the intermediate values were random, what you say would be true, and
interpolation would be a poor choice. But because of the anti-alias
filter, lens PSF, scene content etc., the values are correlated with the
adjacent pixels.

Cheers,
David

From: Alfred Molon on
In article <hrcido$87f$1(a)news.eternal-september.org>, David J Taylor
says...
> If the intermediate values were random, what you say would be true, and
> interpolation would be a poor choice. But because of the anti-alias
> filter, lens PSF, scene content etc., the values are correlated with the
> adjacent pixels.

There may be some correlation which depends on the scene content, but
randomness as well. An example for randomness would be a meadow with
flowers at a distance - no correlation between the red of the flower and
the green of the leaves.
--

Alfred Molon
------------------------------
Olympus E-series DSLRs and micro 4/3 forum at
http://tech.groups.yahoo.com/group/MyOlympus/
http://myolympus.org/ photo sharing site
From: David J Taylor on

"Alfred Molon" <alfred_molon(a)yahoo.com> wrote in message
news:MPG.2644914f1a258a4198c2c0(a)news.supernews.com...
> In article <hrcido$87f$1(a)news.eternal-september.org>, David J Taylor
> says...
>> If the intermediate values were random, what you say would be true, and
>> interpolation would be a poor choice. But because of the anti-alias
>> filter, lens PSF, scene content etc., the values are correlated with
>> the
>> adjacent pixels.
>
> There may be some correlation which depends on the scene content, but
> randomness as well. An example for randomness would be a meadow with
> flowers at a distance - no correlation between the red of the flower and
> the green of the leaves.
> --
>
> Alfred Molon

The amount of correlation will depend on the lens MTF - what the
point-spread function is, and how the anti-alias filter spreads the image.
That's independent of scene content. In the example you cite, any
atmospheric effects over the period of the exposure time would also
contribute to blurring of distant flowers.

When viewing a print with "distant flowers", the eye can't resolve the
colour detail in any case, if by distant you mean "covering a small number
of pixels". I don't think that anyone is saying that Bayer is the perfect
solution - it isn't - but it is one which works very well in practice.

Cheers,
David

From: nospam on
In article <MPG.2644914f1a258a4198c2c0(a)news.supernews.com>, Alfred
Molon <alfred_molon(a)yahoo.com> wrote:

> > If the intermediate values were random, what you say would be true, and
> > interpolation would be a poor choice. But because of the anti-alias
> > filter, lens PSF, scene content etc., the values are correlated with the
> > adjacent pixels.
>
> There may be some correlation which depends on the scene content, but
> randomness as well. An example for randomness would be a meadow with
> flowers at a distance - no correlation between the red of the flower and
> the green of the leaves.

except that the red flower and green leaves will span more than just
one single pixel, so the correct values can be calculated by looking at
the neighboring pixels.
From: Alfred Molon on
In article <290420101341351528%nospam(a)nospam.invalid>, nospam says...

> there is no perhaps about it. the accuracy can be measured. delta-e
> (colour errors) for foveon is higher, or less accurate.
>
> or just look at the images. they're full of weird colour casts,
> including yellow skin, cyan skies, etc. and it's not consistent even
> among multiple foveon cameras. since there's a huge variability, it
> *can't* be more accurate.

The big issue with Bayer is the effective resolution which is way lower
than the nominal one, because a Bayer sensor only captures 1/3 of the
needed information.

> so is bayer. 1/3 the data for an almost perfect reproduction.

No - Bayer is not a data compression method. And reproduction is not
almost perfect.

> > > foveon is the only existing full colour sensor.
> >
> > So? Full colour cameras exist, which is what matters.
>
> but not a full colour sensor.

Wrong. Foveon is a full colour sensor.

<snip>

> many drawbacks for very little gain.

50% more effective resolution is a huge gain.

> interpolation is *not* guessing.

It is, because the value in the middle does not depend on the
neighbouring values.

> > which often are wrong.
>
> what is the error rate? exactly how often is it wrong?

You can see the impact on the resolution: the Foveon 4.6MP performing as
well (resolution-wise) as an 7-8 MP Bayer sensor, according to the
reviews.

> > You cannot calculate the value of a random variable.
>
> it's not random.

The colour of one pixel does not depend on the colour of neighbouring
pixels. And there are scenes with pixel level colour changes (meadow
with flowers at a distance is an example).

In practice, real world scenes often have large areas of (almost)
homogeneous colour, which is why the performance of Bayer sensors is not
as abysmal as the lack of 2/3 of the needed information would imply.

Which is why the effective resolution of a Bayer sensors averages
somewhere arounc 60%-70% of the nominal, which is very good given that
just 1/3 of the needed data is captured.

However there is the huge issue of colour aliasing and the need to use
AA filters tuned to 1/2 of the pixel count resolution, which is why
Bayer images are so soft at the pixel level.
--

Alfred Molon
------------------------------
Olympus E-series DSLRs and micro 4/3 forum at
http://tech.groups.yahoo.com/group/MyOlympus/
http://myolympus.org/ photo sharing site