Increasing filter size of edge detector

Hello,

I am looking at using edge detectors to detect edges in my images but I as wondering if there was a way of increasing the size of the filters? For example increase Sobel or Prewitt from a 3 x 3 array to a 5 x 5 array or bigger. I am looking to see what the effect of a larger window will have on my edge detection.

Thanks, Andrew

Answer 1

[Nitin]

There is a nice thread on stackoverflow which might give you the values of the filter you are looking for.

[Andrew]

Thanks for the response.

I have looked into fspecial to do this but you can't seem to change the size of sobel and prewitt. It also appears to be the same with kosevi that you mentioned. Ideally what I want to do is take the edge detectors in MATLAB and edit the search area (I hope this makes sense).

[Nitin]

Sorry, I have updated my answer. Hopefully this will help

[DGM]

For sake of making things local and elaborating, let's consider Daniel's Sobel-esque filters.

For a 5x5 filter, you can do

r = 2; % filter radius
[x y] = meshgrid(-r:r);
fk = x./(x.^2 + y.^2);
fk(r+1,r+1) = 0

fk = 5×5

-0.2500 -0.2000 0 0.2000 0.2500 -0.4000 -0.5000 0 0.5000 0.4000 -0.5000 -1.0000 0 1.0000 0.5000 -0.4000 -0.5000 0 0.5000 0.4000 -0.2500 -0.2000 0 0.2000 0.2500

You can simply change r to get a different size filter. If you want, you can transpose it.

If you want your output image to stay within a convenient range, you can just prenormalize the kernel depending on how you intend to apply it. If you intend to apply the filter once to get a directional derivative estimate, you can do

fk = fk/-sum(sum(fk(:,1:r))); % for single-pass

... and your image will conveniently stay within [-1 1].

If you intend to do two orthogonal filter passes to calculate the gradient magnitude, you can normalize like so:

fk = fk/(sqrt(5)*r); % for 2-pass

... and your image will stay within [0 1]

... or you can do whatever you want with it as is.

Answer 2

[Image Analyst]

You can try difference of Gaussians or something. Use fspecial to build the Gaussians. DOG filters are common edge filters.