Re: Fuzzy C-Means Usage - Image Segmentation

Dear Ofek:

Thank you for your answer. I understood the difference between "hard"
and "fuzzy/soft" (quoting you :-) segmentation, and I will stick with
the "hard" one (because it is suitable to me).

The second question arises from the fact that in the Step 4 of the demo
(label every pixel in the image using the results from kmeans), it says
that for every object in the input, kmeans returns an index
corresponding to a cluster (named cluster_index), which is used to
label every pixel in the image, The steps to do this are:

pixel_labels = reshape(cluster_idx,nrows,ncols);
imshow(pixel_labels,[]), title('image labeled by cluster index');

This part here is fine: I could show what the demo calls "image labeled
by cluster index" even if I choose four or more colors to segment the
image.

BUT, when I choose more than three colors to segment the image, the
Step 5 of the demo (create images that segment the H&E image by color)
fails. In this step, by using pixel_labels, it is possible to separate
objects in "hestain.png" by color, which will result in three images
(if we use three colors to segment the image). Using nColors = 3, I can
show exactly three images (the same result as I can get from using
kmeans). BUT (again), choosing four colors, my code now is (nColors =
4):

% instead of "segmented_images = cell(1, 3);"
segmented_images = cell(1, 4);
% instead of "rgb_label = repmat(pixel_labels,[1 1 3]);"
rgb_label = repmat(pixel_labels,[1 1 4]);

for k = 1:nColors
color = he;
color(rgb_label ~= k) = 0;
segmented_images{k} = color;
end

Here, the for-loop causes the following error:

??? Attempt to grow array along ambiguous dimension.

Error in ==> color(rgb_label ~= k) = 0;

I noticed that "color = he;" makes color with dimensionallity [X, Y, 3]
and this is causing the error, but I couldn't manage to realise a way
to overcome it. What I am doing wrong here?

I will also try the same algorithm with my grayscale images.

Thanks in advance for any help.

Yours sincerely,

JanKees


Ofek Shilon wrote:
jkvdpoel, sorry i can't give fuller answers now - perhaps later
today. in the meantime:

1. What do you mean with "interpret the fcm results to a 'hard'
segmentation, as opposed to a fuzzy one". What is the difference
between these two ways of segmenting an image?

'hard' segmentation AFAIK is not an official term - what i meant is
some sort of a definite classification of a pixel to any of the K
clusters. that is what you get, e.g., from K means - or any
traditional clustering algorithm. C means can give you a 'soft'
segmentation (again, just a personal terminology): as all fuzzy
algorithms, it gives you a distribution. the pixel is in cluster 1
with probability P1, in cluster 2 with prob P2 etc - this actually
gives you more info than 'hard' clustering, but in typical
segmentation needs you can't really utilize this info (you typically
need just the bottom line - not knowing the details of your
application, of course). what i did in the previous post is choose
the segment with maximum probability out of C-means output for every
fragment, which is in a way forcing the (more expensive) C-means to
behave and output just the K-means output, and i indeed added a
remark questioning the rationale behind such a move.

2. I want, if possible, to do like the demo: show the segmented
images
in a separate figure. With three colors it performs as the original
demo, but the problem arises when I try to segment with more than
three
colors (four colors, for example). I know that in RGB space there
are
only three colors and this is causing the problem, but supose that
I
want to see if there is more than three classes of colors in the
image.
How can I show the other colors in separate figures?


in RGB space there are 256^3 colors, which is quite a bit more than
3. red, green or blue don't have any inherent special status. there
shouldn't be any problems in using the exact same algorithms in 4+
classes - can u specify which 'problem arises' in that case?

3. I also have grayscale images that I want to segment in various
classes (in my application the images are from a CAT scan). Is it
possible to apply the FCM to this kind of images? If so, how can I
do
it to find, for example, five types of different tissues?

this would, in fact, be easier, as your dataset here is 1-
dimensional, as opposed to the 2-dim 'ab' data (extracted from L*a*b
transform) in the demo.

Ofek

.

Relevant Pages

  • Re: HUGE-model pointer math best practice?
    ... segment and the carry bit would be indexing the wrong place in the ... mov ax,;Fetch number of this cluster ... Search for the label getfat12ent. ...
    (comp.lang.asm.x86)
  • Re: File Server Clustering Question
    ... individually will be able to handle all the DFS roots if there is a ... different public network segment. ... It is required that all nodes of the same cluster be on the same network ... MVP - Windows Server - Clustering ...
    (microsoft.public.windows.server.clustering)
  • Re: Disk Drives fail on cluster creation
    ... If the Drives are removed from the cluster as resources, ... >> segment for the quorum and a raid 5 segment for data. ...
    (microsoft.public.windows.server.clustering)
  • Re: Change IP on Windows 2008 cluster
    ... I am assuming if you are moving this cluster to another segment, you'll change the node IPs too. ... Change the IP addresses in the 'Cluster Group' and the file shares group to reflect the new IPs. ...
    (microsoft.public.windows.server.clustering)
  • Re: Bezier lengths w/o lots of resources
    ... (This could be a naive question -- I haven't done the math to see how ... The math everyone is avoiding is for each line segment ... but that's like numerically integrating expstep by step because ... so the algorithm had to handle all cases. ...
    (comp.arch.embedded)