Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
Drexel University or The Math Forum.
|
|
|
Math Forum
»
Discussions
»
sci.math.*
»
sci.math.symbolic.independent
Topic:
[Klicken Sie auf den Stern, um dieses Thema zu verfo
lgen] Online multi-frame blind deconvolution with super-res
olution and saturation correction M. Hirsch, S. Harmeling, S
. Sra, and B. Schölkopf
Replies:
1
Last Post:
Nov 9, 2011 10:31 AM
|
 |
|
|
|
[Klicken Sie auf den Stern, um dieses Thema zu verfo
lgen] Online multi-frame blind deconvolution with super-res
olution and saturation correction M. Hirsch, S. Harmeling, S
. Sra, and B. Schölkopf
Posted:
Nov 9, 2011 8:57 AM
|
|
Hello,
I am currently working on the implementation of the above paper.
Unfortunately there is one piece of important information missing for
me. I don't now the mentioned derivatives in the paper, because I am
not very informed about convolution algebra. Especially, I need "nabla
L", or "nabla l" for the stochastic gradient decent. And I need
dError/
df to compute the argmin expression for the convolution kernel f. I
plan to use gradient decent also in this part of the algortithm.
In the future, I plan to introduce some linear Maps (Matrices) as
Extension to the Algorithm. There is one linear map for performing
downsampling (A) and a geometric warp Matrix (C).
Then the Error expression becomes
|| y - A(f * Cx) ||^2
Should it be possible to compute the derivatives even in this case?
These are d/df, d/dx.
I think constant linear map should be no problem, but I can't compute
it myself.
|
|
|
|
