Find location of cross in image.

I have an image that contains a cross in it. It also also has circles, squares and other features.

How can I detect the location of the cross to sub pixel accuracy?

Matrox MIL has a Geometric Model Finder which works very well. Does Matlab have anything similar?

[KALYAN ACHARJYA]

Can you share a sample image?

Answer 1

[Rajani Mishra]

I have tried below code on an image containing a cross in it, it detects the center location of the cross.

Please refer below for the code:

1. Convert image to grayscale image and then binarization

img = imread('test.jpg');
I = rgb2gray(img);
BW = imbinarize(I,0.25);

2. As the result may have missing parts dilate the image to join the parts

se = strel('line',11,90);
BW2 = imdilate(BW,se);

3. Use morphological ‘shrink’ operation to shrink image to a point

 BW3 = bwmorph(BW2,'shrink',Inf);

4. Find the position of the point

[row,col] = find(BW3);

Hope this helps.

Answer 2

[Daniel]

Hello Hans,

this might be an answer to a long forgotten question. But there may be others out there who are looking for a solution.

Depending on how your image looks this may be complicated to accomplish. However, if you managed to narrow it down to something similar like the 'blobs.png' example, a similar approach like mine might work.

It works because the cross has a big bounding box compared to its filled connected area.

load blobs example image

BW = imread('blobs.png');

Use regionprops to evaluate each individual region using 'FilledArea'

imgprops = regionprops('table',BW,'FilledArea','BoundingBox','Centroid');

The cross is assumed to have a minimum in the filled area to bounding box size ratio (for this image). A diagonal line would have an even lower ratio and the cross will not be the minimum.

[~,k] = min(imgprops.FilledArea ./ prod(imgprops.BoundingBox(:,[3 4]),2));

Plot the result:

figure()
imshow(BW)
hold on
scatter(imgprops.Centroid(k,1),imgprops.Centroid(k,2),'r*');

●

[DGM]

Even if we assume all the blobs are skeletal, the ratio of the filled area to the bb area is scale-dependent. If we discard objects with an Euler characteristic ~=1, then we don't strictly need the filled area for the added discrimination. We can just use 'extent', which is (unfilled blob area)/(bb area). Either way, the ratio is scale-dependent.

inpict = imread('crosses.png');

% some properties
S = regionprops(inpict,'extent','centroid');

% show it
imshow(inpict)
os = [10 10];
for k = 1:numel(S)
    text(S(k).Centroid(1)+os(1),S(k).Centroid(2)+os(2), ...
        num2str(S(k).Extent,2),'color','y')
end

This means that crosses, L-shapes, T-shapes, and straight lines can't be distinguished based on this ratio alone. If we're dealing strictly with skeletal shapes, perhaps we could count endpoints and branchpoints, but that can get tricky, as things don't always skeletonize cleanly.

If we're not dealing with skeletal shapes, then the area ratio is dependent on the thickness of the cross.

[Daniel]

Hello @DGM

Yes, you are absolutely right. Thank you for your answer!

But as there was no image uploaded from Hans, we don't know how accurate and robust the algorithm needs to be. I actually ended up using the "extent" property for my use case as well.

Best regards.