conv2 with NaNs?

[Temu Gautama]

Hi guys,

Does anybody have a 2D convolution function that can handle NaNs (the
NaNs should not be taken into account into the convolution sum)?
It's fairly straightforward to code, but I was wondering if there's
an efficient implementation around. An example problem:

>> A = randn(100,100);
>> INDr=randperm(10000);A(INDr(1:100))=NaN;
>> pcolor(A); shading flat
>> [x y]=meshgrid(-2:4/24:2);H=exp(-(x.^2+y.^2)./2);
>> pcolor(conv2(A,H));shading flat

Most of the convolution contains NaNs, since it is set to NaN if at
least one of the values within the region (size(H)) is NaN.

???

Temu

[herbert(DOT)ramoser]

Does setting all NaN regions to zero solve your problem? If not what
should happen with the NaNs in the convolution?

-Herbert

[Alois Schloegl]

Temu Gautama wrote:

Try the attached function

and see if this is what you expect
pcolor(conv2nan(A,H));shading flat

Alois

[Temu Gautama]

Hi Herbert,

> Does setting all NaN regions to zero solve your problem? If not
> what
> should happen with the NaNs in the convolution?

You're absolutely right! I was a bit confused (was thinking about
mean, rather than sum).

Thanks!

Temu

[Temu Gautama]

Alois, I didn't see an attachment, there?

[Tim Zaman]

Alois Schloegl <a.sch...@ieee.org> wrote in message <3FE81C24...@ieee.org>...
>
> (...)

>
>
> Try the attached function
>
> and see if this is what you expect
> pcolor(conv2nan(A,H));shading flat
>
>
> Alois
>
>

> (...)


Alois, this is EXCELLENT!