Color segmentation using kmeans clustering

I shared before kmeans Algorithm （ Portal ： Data mining algorithms —K-Means Algorithm ）, In this issue, let's share Kmeans Clustering realizes color segmentation , Use L*a*b* Sum of color Spaces K Mean clustering automatic color segmentation .

step 1： Read images

Read hestain.png,

he = imread('hestain.png');
imshow(he), title('H&E image');

Copy

step 2： Take the image from RGB The color space is converted to L*a*b* Color space

L*a*b* Color space （ Also known as CIELAB or CIE L*a*b*） Be able to quantify visual differences .

L*a*b* The color space is from CIE XYZ Trichromatic value derived .L*a*b* Space contains photometric layers 'L*'、 Chrominance layer 'a*'（ Indicates that the color falls along the red - The position of the green axis ） And chroma layer 'b*'（ Indicates that the color falls along the blue - The position of the Yellow axis ）. All color information is 'a*' and 'b*' layer . Euclidean distance measurement can be used to measure the difference between two colors .

Use rgb2lab Convert image to L*a*b* Color space .

lab_he = rgb2lab(he);

Copy

step 3： use K Mean clustering pairs are based on 'a*b*' Space color classification

Clustering is a method of separating groups of objects .K Mean clustering treats each object as having a position in space . It divides objects into partitions , Make the objects in each cluster as close to each other as possible , And as far away from objects in other clusters as possible .K Mean clustering requires you to specify the number of clusters to divide and the distance metric used to quantify the distance between two objects .

Because the color information is based on 'a*b*' Color space , So your object has 'a*' and 'b*' Value in pixels . Convert data to data type single, So as to meet with imsegkmeans Use a combination of . Use imsegkmeans Cluster objects into three clusters .

ab = lab_he(:,:,2:3);
ab = im2single(ab);
nColors = 3;
%  Repeat clustering three times , Avoid local optimizations 
pixel_labels = imsegkmeans(ab,nColors,'NumAttempts',3);

Copy

For each object in the input ,imsegkmeans An index or label corresponding to the cluster will be returned . Label each pixel in the image with the label of the pixel .

imshow(pixel_labels,[])
title('Image Labeled by Cluster Index');

Copy

step 4： Create a color split H&E The image of the image

Use pixel_labels, Can be separated by color hestain.png Objects in the , This will produce three images .

mask1 = pixel_labels==1;
cluster1 = he .* uint8(mask1);
imshow(cluster1)
title('Objects in Cluster 1');

Copy

mask2 = pixel_labels==2;
cluster2 = he .* uint8(mask2);
imshow(cluster2)
title('Objects in Cluster 2');

Copy

mask3 = pixel_labels==3;
cluster3 = he .* uint8(mask3);
imshow(cluster3)
title('Objects in Cluster 3');

Copy

step 5： Split core

cluster 3 Contains blue objects . Please note that , There are dark blue and light blue objects . You can use L*a*b* In color space 'L*' Layer to separate dark blue and light blue . The nucleus is dark blue .

As mentioned earlier ,'L*' The layer contains the brightness value of each color . Extract the brightness value of pixels in this cluster , And use imbinarize Set the threshold with the global threshold . A mask is_light_blue Gives the index of light blue pixels .

L = lab_he(:,:,1);
L_blue = L .* double(mask3);
L_blue = rescale(L_blue);
idx_light_blue = imbinarize(nonzeros(L_blue));

Copy

Copy the mask of the blue object mask3, Then remove the light blue pixels from the mask . Apply the new mask to the original image and display the result . Only dark blue nuclei are visible .

blue_idx = find(mask3);
mask_dark_blue = mask3;
mask_dark_blue(blue_idx(idx_light_blue)) = 0;

blue_nuclei = he .* uint8(mask_dark_blue);
imshow(blue_nuclei)
title('Blue Nuclei');

Copy