Re: Rookie having problems with some filter code. Any help?

Don Bruder <dakidd@xxxxxxxxx> writes:

>> If you're just going to write out the image again, negative values
>> should be clamped to zero, and values greater than 255 should be clamped
>> to 255 - not wrapped around to zero.

>Hmmm... So simple clamping is, despite the counterintuitiveness
>involved, the way to go? Without chasing every possiblity, it seems
>likely to me that would end up with a whole slew of pixels whose
>color-component values are pegged at either 0 or 255, with comparatively
>little "middle ground".

It depends on the filter, and what you expect of it. For most filters
applied to photographic images, the output is supposed to look rather
like the input. Sharpening filters *can* cause overrange values, either
less than zero or greater than 255, but the right thing to do is to
clamp to the nearest representable value. So "blacker than black"
(negative) values become black (zero), and "super white" (>255) values
become white. That's better than turning black into white or vice
versa, which is what happens if you take the result mod 255.

>OK, that does make at least SOME sense, but it still hits the same wall
>mentioned above - At first glance, it would seem that clamping,
>particularly in a "many-control-one" situation needing many
>multiplications (some of them possibly involving negative multipliers)
>and additions to arrive at a value, whether it happens "in the middle"
>or "at the end" of an ops sequence, would tend to skew things pretty
>strongly to the extremes of the range, eating the bejeebers out of the
>"middle" values in the process.

Only if you're doing something that grossly exaggerates the size of
differences, such as an edge-finding filter.

>> There are occasions (e.g. differencing) when keeping the result mod 256
>> is the right thing to do - but that's not usually the case.

>Are you saying "differencing" to mean what I think you are? As in "XOR",
>subtract, or otherwise combine two frames (or segments of frames) -
>let's call them "F1" and "F2", and the "combined" version that results
>"R" for convenience - so that if a pair of pixels, F1[x,y] and F2[x,y],
>share a common RGB value - let's say they're both r=101, g=30, b=222 -
>then R[x,y]'s color will be set to <some arbitrary color value>? (which
>will probably, but not necessarily, be white or black due to common
>practice, but whatever it actually is, the intent is to convey a
>"nothing to see here" message.)

One way of showing differences is to subtract two images pixel by pixel,
component by component, and then optionally add 128. Then differences
show up as deviations from grey.

But what I meant is this: Suppose you have two values A and B, and
calculate the difference D in ordinary math: D = A - B. Then you can
reconstruct either of the originals from the other plus the difference:
A = B + D, B = A - D. This is *also* true if A, B, and D are images,
provided *no clampling* happens. Clamping loses information, and makes
the operation non-reversible.

In practical terms, this means that D needs to be 9 bits wide, if A and
B are 8 bits wide, and you use ordinary integer arithmetic. However, if
you use *modular* arithmetic, with all results taken mod 256, then all
the expressions above remain true with D only 8 bits wide.

So, using mod-256 arithmetic (which is what CPUs natively provide)
without clamping allows a difference operation that's invertible,
without needing extra bits. It's more difficult to interpret visually
than the usual clamped version though.

Dave
.

Relevant Pages

  • Daily Report # 4347
    ... High-fidelity multi-color HST images are required ... Blue hook stars are a class of hot subluminous horizontal branch ... axi-symmetric orbit superposition code. ... Evaluation of Universal Kalman Filter performance continued. ...
    (sci.astro.hubble)
  • Re: Interpolation And Low-Pass filtering
    ... >>4) Lowpass filter by multiplying in frequency. ... > This causes no problem because we are padding zero. ... >>frequency domain affect is to convolve the spectrum>with a sinc. ...
    (comp.dsp)
  • Re: Digital sensors Nikon Canon
    ... to do less computer doctoring with my Fujifilm pictures from the S5000 ... But you say there's a filter on there too. ... So if you find the images from the fuji better, ... All this assumes you shoot jpeg. ...
    (rec.photo.digital)
  • Re: NASAs Mars probe: NO cmos, NO Bayer
    ... 'Images are recorded by a charge-coupled device similar to those ... There are five cameras all together on Phoenix, though "imager" is ... Each sensor has a filter wheel containing twelve separate ...
    (rec.photo.digital.slr-systems)
  • Re: Desired Sum of Two Curves; suggestions?
    ... > I'm working on an adaptive digital filter that requires two ... The sum b+cresults in a curve that rises from zero at ... squared error of Ea and the squared error of Eb. ...
    (sci.math)
Quantcast