Archived from groups: alt.video.laserdisc (
More info?)
In article <46882a5be3f335ad36ac936469ba6f60@localhost.talkaboutvideo.com>,
"half_eaten" <half_eaten@nospam.yahoo.com> writes:
> I was thinking of programming my own software to attempt to take care of
> this.. could average the frame to the left and right of that pixel with
> the pixel in the middle and mix them together at a ratio determined by the
> color strength of that pixel.
>
> for example if there the color isn't very strong, it won't blend much at
> all, but if it's say a VERY bright red or blue it would mix the pixels
> together quite a bit. this should smooth out chroma noise. plus since LDs
> are only 425 pixels horizontal resolution this wouldn't cause too much
> loss of sharpness.
>
You are considering an ah-hoc (not necessarily bad) nonlinear digital
filter. If you have had some training in digital filtering, I'd suggest
doing a little more research, and also doing some 3D filtering (doing
some frame averaging.) All of the filtering techniques should likely
take advantage of some nonlinear processing (changing the amount of
filtering or averaging based upon signal characteristics.)
You can do a reasonably good job without a theoretical background.
If you use several techniques, you might get a 3dB (or maybe more)
noise reduction. If you use several high power techniques, you might
be able to get alot more... When I use lots of NR (both SW and HW),
I can do a capture that looks ALMOST like a 'soft' DVD from an
LD... (The resolution is like a very good LD, and the SNR is FAIRLY
CLOSE to a moderately good DVD, with very good detail but very low
HF noise.)
Being more specific, what you are doing is doing 2D filtering, and the
video signal is already significantly filtered in the HORIZONTAL
direction, but there is some potential benefit in the vertical direction
(especially if you do the averaging between fields -- directly adjacent
lines, and mitigate SOME of the loss in vertical resolution.) This entails
a delay that is a full field (because adjacent interlaced lines are actually
the full field time away.) If the two fields are very different, then they
shouldn't be averaged. If the difference is between fields minimal, then you
can do significant amounts of averaging without too much damage.
Avoid the RGB domain for most of your filtering and look towards the
Y-Cb-Cr type domain or HVS (if you are a little more aggressive.) Frankly,
you might be able to take advantage of working in both domains.
John
.
Facebook
Twitter