when to use "conv2", when to use "fft"->"multiplication"->"ifft"?

Hi all,

I am doing 2D image filtering with sometimes filter size larger than image 
size.

I am wondering which of the following scheme is the fastest, using Matlab:

1) conv2, 2D convolution directly;
2) conv2, but decompose the filter into two separable filters and then conv2 
will first convolve along X axis and then convolve along Y axis;
3) fft2, multiplication, ifft2, all in 2D domain;
4) fft, multiplication, ifft, all in 1D case and work on convolutions of 
separable filters along X axis and Y axis; (similar to 2, but do 1D 
convolution all in frequency domain).

Images and filters are typically hundreds and thousands of pixels, is FFT 
method always faster? 2D FFT or use 1D separable FFT?

Thanks

Why not just trying and pick the fastest one for you ?

For 2D FFT, and 1D separable FFT ... if you are using Matlab,
I think 2D FFT is a little faster.  If you are using C library, the
performance should be very very close.

Regards,
KC

>I am wondering which of the following scheme is the fastest, using Matlab:
>
>1) conv2, 2D convolution directly;
>2) conv2, but decompose the filter into two separable filters and then conv2 
>will first convolve along X axis and then convolve along Y axis;
>3) fft2, multiplication, ifft2, all in 2D domain;
>4) fft, multiplication, ifft, all in 1D case and work on convolutions of 
>separable filters along X axis and Y axis; (similar to 2, but do 1D 
>convolution all in frequency domain).


AFIK, Matlab's conv2 uses FFT/Multiply/IFFT when sizes are powers of
2.  Don't quote me on that though.  You should find conv2 is about the
same speed as the FFT route.

A