I have a 3D image, divided into contiguous regions where each voxel has the same value. The value assigned to this region is unique to the region and serves as a label. The example image below describes the 2D case:
1 1 1 1 2 2 2
1 1 1 2 2 2 3
Im = 1 4 1 2 2 3 3
4 4 4 4 3 3 3
4 4 4 4 3 3 3
I want to create a graph describing adjaciency between these regions. In the above case, this would be:
0 1 0 1
A = 1 0 1 1
0 1 0 1
1 1 1 0
I'm looking for a speedy solution to do this for large 3D images in MATLAB. I came up with a solution that iterates over all regions, which takes 0.05s per iteration - unfortunately, this will take over half an hour for an image with 32'000 regions. Does anybody now a more elegant way of doing this? I'm posting the current algorithm below:
labels = unique(Im); % assuming labels go continuously from 1 to N
A = zeros(labels);
for ii=labels
% border mask to find neighbourhood
dil = imdilate( Im==ii, ones(3,3,3) );
border = dil - (Im==ii);
neighLabels = unique( Im(border>0) );
A(ii,neighLabels) = 1;
end
imdilate is the bottleneck I would like to avoid.
Thank you for your help!
I came up with a solution which is a combination of Divakar's and teng's answers, as well as my own modifications and I generalised it to the 2D or 3D case.
To make it more efficient, I should probably pre-allocate the r and c, but in the meantime, this is the runtime:
For a 3D image of dimension 117x159x126 and 32000 separate regions: 0.79s
For the above 2D example: 0.004671s with this solution, 0.002136s with Divakar's solution, 0.03995s with teng's solution.
I haven't tried extending the winner (Divakar) to the 3D case, though!
noDims = length(size(Im));
validim = ones(size(Im))>0;
labels = unique(Im);
if noDims == 3
Im = padarray(Im,[1 1 1],'replicate', 'post');
shifts = {[-1 0 0] [0 -1 0] [0 0 -1]};
elseif noDims == 2
Im = padarray(Im,[1 1],'replicate', 'post');
shifts = {[-1 0] [0 -1]};
end
% get value of the neighbors for each pixel
% by shifting the image in each direction
r=[]; c=[];
for i = 1:numel(shifts)
tmp = circshift(Im,shifts{i});
r = [r ; Im(validim)];
c = [c ; tmp(validim)];
end
A = sparse(r,c,ones(size(r)), numel(labels), numel(labels) );
% make symmetric, delete diagonal
A = (A+A')>0;
A(1:size(A,1)+1:end)=0;
Thanks for the help!
Try this out -
Im = padarray(Im,[1 1],'replicate');
labels = unique(Im);
box1 = [-size(Im,1)-1 -size(Im,1) -size(Im,1)+1 -1 1 size(Im,1)-1 size(Im,1) size(Im,1)+1];
mat1 = NaN(numel(labels),numel(labels));
for k2=1:numel(labels)
a1 = find(Im==k2);
for k1=1:numel(labels)
a2 = find(Im==k1);
t1 = bsxfun(#plus,a1,box1);
t2 = bsxfun(#eq,t1,permute(a2,[3 2 1]));
mat1(k2,k1) = any(t2(:));
end
end
mat1(1:size(mat1,1)+1:end)=0;
If it works for you, share with us the runtimes as comparison? Would love to see if the coffee brews any faster than half an hour!
Below is my attempt.
Im = [1 1 1 1 2 2 2;
1 1 1 2 2 2 3;
1 4 1 2 2 3 3;
4 4 4 4 3 3 3;
4 4 4 4 3 3 3];
% mark the borders
validim = zeros(size(Im));
validim(2:end-1,2:end-1) = 1;
% get value of the 4-neighbors for each pixel
% by shifting the images 4 times in each direction
numNeighbors = 4;
adj = zeros([prod(size(Im)),numNeighbors]);
shifts = {[0 1] [0 -1] [1 0] [-1 0]};
for i = 1:numNeighbors
tmp = circshift(Im,shifts{i});
tmp(validim == 0) = nan;
adj(:,i) = tmp(:);
end
% mark neighbors where it does not eq Im
imDuplicates = repmat(Im(:),[1 numNeighbors]);
nonequals = adj ~= imDuplicates;
% neglect the border
nonequals(isnan(adj)) = 0;
% get these neighbor values and the corresponding Im value
compared = [imDuplicates(nonequals == 1) adj(nonequals == 1)];
% construct your 'A' % possibly could be more optimized here.
labels = unique(Im);
A = zeros(numel(labels));
for i = 1:size(compared,1)
A(compared(i,1),compared(i,2)) = 1;
end
#Lisa
Yours reasoning is elegant, though it obviously gives wrong answers for labels on the edges.
Try this simple label matrix:
Im =
1 2 2
3 3 3
3 4 4
The resulting adjacency matrix , according to your code is:
A =
0 1 1 0
1 0 1 1
1 1 0 1
0 1 1 0
which claims an adjacency between labels "2" and "4": obviously wrong. This happens simply because you are reading padded Im labels based on "validim" indices, which now doesn't match the new Im and goes all the way down to the lower borders.
Related
I have got the following code. I need to rewrite it without looping. How should I do it?
l1 = [1 2 3 2 1];
l2 = [3 4 4 5 4];
A = zeros(5,5);
for i=1:5
A(i, l1(i):l2(i)) = 1;
end
A
You can use bsxfun -
I = 1:5 % Array corresponding to iterator : "for i=1:5"
out = bsxfun(#le,l1(:),I) & bsxfun(#ge,l2(:),I)
If you need a double datatype array, convert to double, like so -
out_double = double(out)
Add one more into the mix then! This one simply uses a cumsum to generate all the 1s - so it does not use the : operator at all - It's also fully parallel :D
l1 = [1 2 3 2 1];
l2 = [3 4 4 5 4];
A = zeros(5,5);
L1 = l1+(1:5)*5-5; %Convert to matrix location index
L2 = l2+(1:5)*5-5; %Convert to matrix location index
A(L1) = 1; %Place 1 in that location
A(L2) = 1; %Place 1 in that location
B = cumsum(A,1) ==1 ; %So fast
Answer = (A|B)'; %Lightning fast
Answer =
1 1 1 0 0
0 1 1 1 0
0 0 1 1 0
0 1 1 1 1
1 1 1 1 0
Here is how you could build the matrix without using a loop.
% Our starting values
l1 = [1 2 3 2 1];
l2 = [3 4 4 5 4];
% Coordinate grid of the right size (we don't need r, but I keep it there for illustration)
[r,c] = ndgrid(1:5);
% Build the logical index based on our lower and upper bounds on the column indices
idx_l1=bsxfun(#ge,c,l1');
idx_l2=bsxfun(#le,c,l2');
% The result
A = zeros(size(idx_l1));
A(idx_l1&idx_l2)=1
You may need something like [r,c] = ndgrid(1:numel(l1),1:10).
Also if your matrix size is truly huge and memory becomes an issue, you may want to stick to a loop anyway, but for 'normal size' this could be faster.
There should be some skepticism in every vectorization. If you measure the time actually your loop is faster than the given answers, mostly because you only perform in place write.
Here is another one that would probably get faster for larger sizes but I haven't tested:
tic
myind = [];
for i = 1:5
myind = [myind (5*(i-1))+[l1(i):l2(i)]];
end
A(myind) = 1;
toc
gives the transposed A because of the linear indexing order.
I'm performing input-output calculations in Octave. I have several matrices/vectors in the formula:
F = f' * (I-A)^-1 * Y
All vectors probably contain zeroes. I would like to omit them from the calculation and just return 0 instead. Any help would be greatly appreciated!
Miranda
What do you mean when you say "omit them"?
If you want to remove zeros from a vector you can do this:
octave:1> x=[1,2,0,3,4,0,5];
octave:2> x(find(x==0))=[]
x =
1 2 3 4 5
The logic is: x==0 will test each element of x (in this case the test is if it equals zero) and will return a vector of 0's and 1's (0 if the test is false for that element and 1 otherwise)
So:
octave:1> x=[1,2,0,3,4,0,5];
octave:2> x==0
ans =
0 0 1 0 0 1 0
The find() function will return the index value of any non-zero element of it's argument, hence:
octave:3> find(x==0)
ans =
3 6
And then you are just indexing and removing when you do something like:
octave:5> x([3, 6]) = []
x =
1 2 3 4 5
But instead you do it with the output of the find() function (which is the vector [3,6] in this case)
You can do the same for matrices:
octave:7> A = [1,2,0;4,5,0]
A =
1 2 0
4 5 0
octave:8> A(find(A==0))=[]
A =
1
4
2
5
Then use the reshape() function to turn it back into a matrix.
I want to change the individual elements in a matrix by different values simultaneously.
How do I do that?
For example: I want to change the first element in matrix A by certain amount and the second element by a different amount simultaneously.
{ A = [1; 2]
% instead of doing A(1) = .....
A(2) = .....
}
You can access the elements of a vector or matrix and replace them.
For a vector this is intuitive.
octave:16> A = 1:9
A =
1 2 3 4 5 6 7 8 9
octave:17> A([1 3 5 7 9]) = 0
A =
0 2 0 4 0 6 0 8 0
This can be done for a matrix as well. The elements of a matrix are arranged in a column-first manner. You can use a single index to access the elements of a matrix.
octave:18> A = [1 2 3; 4 5 6; 7 8 9]
A =
1 2 3
4 5 6
7 8 9
The 2nd element of A is the same as A(2, 1). The 4th element of A is the same as A(1, 2).
octave:21> A(2)
ans = 4
octave:22> A(4)
ans = 2
So, you can set all the odd elements of A to 0 in one go like this:
octave:19> A([1 3 5 7 9]) = 0
A =
0 2 0
4 0 6
0 8 0
Just add a vector with the differences. A += [0.1; 0.2]
octave:1> A = [1; 2];
octave:2> A += [0.1; 0.2]
A =
1.1000
2.2000
There are two vectors:
a = 1:5;
b = 1:2;
in order to find all combinations of these two vectors, I am using the following piece of code:
[A,B] = meshgrid(a,b);
C = cat(2,A',B');
D = reshape(C,[],2);
the result includes all the combinations:
D =
1 1
2 1
3 1
4 1
5 1
1 2
2 2
3 2
4 2
5 2
now the questions:
1- I want to decrease the number of operations to improve the performance for vectors with bigger size. Is there any single function in MATLAB that is doing this?
2- In the case that the number of vectors is more than 2, the meshgrid function cannot be used and has to be replaced with for loops. What is a better solution?
For greater than 2 dimensions, use ndgrid:
>> a = 1:2; b = 1:3; c = 1:2;
>> [A,B,C] = ndgrid(a,b,c);
>> D = [A(:) B(:) C(:)]
D =
1 1 1
2 1 1
1 2 1
2 2 1
1 3 1
2 3 1
1 1 2
2 1 2
1 2 2
2 2 2
1 3 2
2 3 2
Note that ndgrid expects (rows,cols,...) rather than (x,y).
This can be generalized to N dimensions (see here and here):
params = {a,b,c};
vecs = cell(numel(params),1);
[vecs{:}] = ndgrid(params{:});
D = reshape(cat(numel(vecs)+1,vecs{:}),[],numel(vecs));
Also, as described in Robert P.'s answer and here too, kron can also be useful for replicating values (indexes) in this way.
If you have the neural network toolbox, also have a look at combvec, as demonstrated here.
One way would be to combine repmat and the Kronecker tensor product like this:
[repmat(a,size(b)); kron(b,ones(size(a)))]'
ans =
1 1
2 1
3 1
4 1
5 1
1 2
2 2
3 2
4 2
5 2
This can be scaled to more dimensions this way:
a = 1:3;
b = 1:3;
c = 1:3;
x = [repmat(a,1,numel(b)*numel(c)); ...
repmat(kron(b,ones(1,numel(a))),1,numel(c)); ...
kron(c,ones(1,numel(a)*numel(b)))]'
There is a logic! First: simply repeat the first vector. Secondly: Use the tensor product with the dimension of the first vector and repeat it. Third: Use the tensor product with the dimension of (first x second) and repeat (in this case there is not fourth, so no repeat.
I here by post the code why I came across while exploring one technique.
Y = repmat((1:m)', [1 n]);
X = repmat(1:n, [m 1]) - labels_left;
X(X<1) = 1;
indices = sub2ind([m,n],Y,X);
final_labels = labels_left;
final_labels(abs(labels_left - labels_right(indices))>=1) = -1;
In above code labels left is single channel image.[m n] is the size of that image. I want to know how this sub2ind works in above code.And Iam also facing problem in the last statement which contains
labels_right(indices)
what the above expression evaluates to.Here labels right is also an image
Maybe a smaller example could help understand:
%# image matrix
M = rand(4,3)
[m n] = size(M)
%# meshgrid, and convert to linear indices
[X,Y] = meshgrid(1:n,1:m)
indices = sub2ind([m,n],Y,X)
%# extract those elements
M(indices)
The matrix M:
>> M
M =
0.95717 0.42176 0.65574
0.48538 0.91574 0.035712
0.80028 0.79221 0.84913
0.14189 0.95949 0.93399
the grid of (x,y) coordinates of all points:
>> X,Y
X =
1 2 3
1 2 3
1 2 3
1 2 3
Y =
1 1 1
2 2 2
3 3 3
4 4 4
converted to linear indices:
>> indices
indices =
1 5 9
2 6 10
3 7 11
4 8 12
then we index into the matrix using those indices.
>> M(indices)
ans =
0.95717 0.42176 0.65574
0.48538 0.91574 0.035712
0.80028 0.79221 0.84913
0.14189 0.95949 0.93399
Note that: M(indices(i,j)) = M(Y(i,j)),X(i,j)).