Large Convolution Kernel?

[zveris]

I need to use hilbert transform on images. but it is rather large kernel, while cop's default kernel is 9 x 9. Is there a way to get a larger kernel (maybe with wex) without getting into sdk?

Thanks

Zveris

There was no start for his talents.... (T.Pratchett)

[edward]

Without knowing anything about hilbert transforms, fixed size kernels are doable in VEX. It's a bit annoying though since it has doesn't support arrays natively.

[zveris]

well, i need something like that:

-h4 0 -h3 0 -h2 0 -h1 0 h1 0 h2 0 h3 0 h4

it is simple single row, but 15 column kernel.

How it would be possible to do it in vex?

Thanks!

[edward]

Sure. Again, I don't know what type of functions you need.

[malexander]

The convolve COP also accepts an image as a second input for the kernel data. If you could make a 15x1 image of your kernel (in FP, likely, since it has negative components) , then you could feed it into the convolve COP to do the job. Even if you had to generate the kernel through a VEX generator, the convolve COP / kernel generator COP combination would likely be much faster than the 15 independent lookups that the Hilbert convolve VEX cop would need to perform. You could create a digital asset out of the small network to create a Hilbert Transform COP, if you needed this operation frequently.