Comments on: A structuring element decomposition “gotcha”

By: Steve

Steve — Thu, 07 Oct 2010 13:00:48 +0000

Cris—I see. My fault, I suppose. I’m the one who designed and implemented strel.

By: Cris Luengo

Cris Luengo — Thu, 07 Oct 2010 12:52:32 +0000

Steve, this is what I'm referring to:

imshow( (fspecial('disk',100)>0) - getnhood(strel('disk',100,0)) )

I thought it was an interesting find... good to be aware of.

By: Steve

Steve — Wed, 06 Oct 2010 12:12:39 +0000

Cris—The variation in size for strel(‘disk’,5) comes from the structuring element decomposition. There isn’t an exact decomposition for a disk-shaped structuring element, so this call is producing an approximation. Use strel(‘disk’,5,0) to avoid the approximation.

By: Cris Luengo

Cris Luengo — Wed, 06 Oct 2010 09:11:05 +0000

@ Ken’s post:

1- Why are these the same size? I would have hoped for a little more consistency across functions in one toolbox! :)
disk = strel(‘disk’,5);
disk = fspecial(‘disk’,4);

2- The various operations compared in this post do not do the same thing. Replace ‘a’ with ‘a=imread(‘moon.tif’)’ or something similar, and you’ll see. ‘imdilate’ is not limited to binary images.

3- The timing difference drops significantly with this larger input image (moon.tif). And ‘imdilate’ turns out to be faster if you make the disk with a radius of 15 instead of 5 (this is where the decomposition of the structuring element becomes significant). Don’t know where I’m heading with this point, I guess most of the differences is in the overhead rather than the computation.

By: Steve

Steve — Mon, 01 Jun 2009 13:43:40 +0000

Alex—The appropriate structuring element shape depends on your application and data set, just like choosing between a lowpass filter and a highpass filter. You might want to get a good textbook on image processing.

By: Alex

Alex — Sun, 31 May 2009 17:37:22 +0000

Hi.. I cannot understand what is written in matlab help. Could you give me a brief description on how to select the appropriate structuring element parameter? wheter is disc or rectangle or wad.

By: Steve

Steve — Fri, 19 Dec 2008 21:08:29 +0000

Ken—I get similar results. Clearly, we have a lot of room to improve our dilation implementation! We’re looking into it.

By: Steve

Steve — Sat, 13 Dec 2008 16:59:08 +0000

Oliver—I didn't say bwlabel allows arbitrary strels, just arbitrary connectivities. I would be interested to hear more about your application, if you care to share it. For example, when would it make sense to say that the two 1-valued pixels below are connected?

1 0 0 0 0 0 0 0 1

Anyway, some of the material in my series on connected component labeling algorithms might be useful to you.

By: Oliver Woodford

Oliver Woodford — Sat, 13 Dec 2008 16:35:47 +0000

Thanks, Steve. However, the CONN matrix doesn’t allow arbitrary strel neighbourhoods. In particular, the neighbourhoods are limited to 3×3 (for 2d) and must be symmetric. For the problem I have (with a 9×9 neighbourhood), BWLABELN can’t be used.

By: Steve

Steve — Sat, 13 Dec 2008 16:02:08 +0000

Oliver—bwlabeln supports labeling operations with an arbitrary connectivity definition. See the doc regarding the CONN input argument. You can use bwlabeln on two-dimensional inputs.