Poutnik
27th December 2005, 12:02
Hello,
I process almost exclusive air TV analog captures with high to low noise quality.
LRemoveDust is my favourite denoiser.
I think about an easy method how to choose proper limit values
for LRemoveDust script (or threshold values for temporal denoisers as well).
I would like to know,
what you think about my ideas,
how you do such desicions (if any non visual)
and if you have any ready to use numerical solution.
Here are my ideas....
----------------------
1) Visual compare - supposed safe and efficient, but...
For me is often difficult to justify low changes in noise level and detail quality and my eyes are getting tired.
----------------------
2) Compression test
Trying several LRemoveDust limit values
Performing XviD compression check with fix quant
Needs extra work. It can be difficult to realize
if compression gain is done by denoising or detail removing.
OTOH I think skilled person can choose proper limit on compression curve shape.
E.g. one can choose highest limit value with significant compression gain,
or even math fans can do some math analysis of such function (nothing for i4004 :-) ).
----------------------
Next 2 ideas are based on simple statistical estimation...
Supposing noise has normal distribution.
I can see relation between noise sigma and average absolute difference (AAD?) between previous and next still frames.
Lets say noise level sigma = s.
Sigma of noise of frame difference is sd = sqrt(2)*s (s^2=s1^2+s2^2)
AVG absolute value of normal distributian is 2/3 of sigma (excel numerical result)., i.e. 2/3 of sd, or cca s (1.4 * 2/3*s)
Example:
if limit = 2 * sd (level 95%) = cca 3 * AAD,
we can expect 95 % of noise in still video can be eliminated and rest is limited.
We can use different koef for another removing level.
For changing frames it is more complicated, but I think "min AAD" could
be considered as good noise measure.
So I can estimate appropriate limit value, based on AAD value.
-----------------------------
3) PeachSmoother Baseline a/o Noiselevel values.
The question is, how they are related to AAD ?
( Numerical comparing not yet done)
-----------------------------
4) Runtime functions for Y/u/V differences
Use low motion/schange part of video if posssible.
Use minimal AAD value ( supposed caused by noise only).
Use AAD and proper limit value relation .
I process almost exclusive air TV analog captures with high to low noise quality.
LRemoveDust is my favourite denoiser.
I think about an easy method how to choose proper limit values
for LRemoveDust script (or threshold values for temporal denoisers as well).
I would like to know,
what you think about my ideas,
how you do such desicions (if any non visual)
and if you have any ready to use numerical solution.
Here are my ideas....
----------------------
1) Visual compare - supposed safe and efficient, but...
For me is often difficult to justify low changes in noise level and detail quality and my eyes are getting tired.
----------------------
2) Compression test
Trying several LRemoveDust limit values
Performing XviD compression check with fix quant
Needs extra work. It can be difficult to realize
if compression gain is done by denoising or detail removing.
OTOH I think skilled person can choose proper limit on compression curve shape.
E.g. one can choose highest limit value with significant compression gain,
or even math fans can do some math analysis of such function (nothing for i4004 :-) ).
----------------------
Next 2 ideas are based on simple statistical estimation...
Supposing noise has normal distribution.
I can see relation between noise sigma and average absolute difference (AAD?) between previous and next still frames.
Lets say noise level sigma = s.
Sigma of noise of frame difference is sd = sqrt(2)*s (s^2=s1^2+s2^2)
AVG absolute value of normal distributian is 2/3 of sigma (excel numerical result)., i.e. 2/3 of sd, or cca s (1.4 * 2/3*s)
Example:
if limit = 2 * sd (level 95%) = cca 3 * AAD,
we can expect 95 % of noise in still video can be eliminated and rest is limited.
We can use different koef for another removing level.
For changing frames it is more complicated, but I think "min AAD" could
be considered as good noise measure.
So I can estimate appropriate limit value, based on AAD value.
-----------------------------
3) PeachSmoother Baseline a/o Noiselevel values.
The question is, how they are related to AAD ?
( Numerical comparing not yet done)
-----------------------------
4) Runtime functions for Y/u/V differences
Use low motion/schange part of video if posssible.
Use minimal AAD value ( supposed caused by noise only).
Use AAD and proper limit value relation .