imhmin

Create a 10-by-10 sample image. Add three regional minima, each consisting of an area of connected pixels surrounded by higher intensity values.

a = 10*ones(10,10);
a(2:4,2:4) = 7;  
a(6:8,6:8) = 2;
a(1:3,7:9) = 13;
a(2,8) = 10;

This image is a grayscale representation of the pixel values. The depth of each minimum depends on the surrounding pixel values.

Apply the H-minima transform that decreases the depth of regional minima by up to 4.

h = 4;
b = imhmin(a,h);

This image is a grayscale representation of the transformed image. The transform fully suppresses two of the minima. The transform partially suppresses the deepest minimum, and adds 4 to the intensity values of the pixels in that minimum.

You can suppress shallow regional minima to avoid oversegmentation during watershed segmentation.

Load an RGB image of pears to segment. Convert the image to grayscale and display it. The center of each pear is bright, corresponding to a regional maximum.

RGB = imread("pears.png");
I = im2gray(RGB);
imshow(I)

In watershed segmentation, the image is analogous to a surface comprised of watershed lines and catchment basins. When water flows into the surface, it pools in the catchment basins. In a grayscale image, local minima are the catchment basins. To segment the pears, invert the image so the centers of the pears become regional minima.

Icomp = imcomplement(I);
imshow(Icomp)

Display the inverted image as a 3-D surface, in which the third dimension for each pixel is its intensity value. The deeper regions for each pear have spiky bottoms, indicating many shallow regional minima, like catchment basins into which water can pool.

surf(Icomp,EdgeColor="none")
colormap(gray)

Segment the unfiltered image and display the result as a label overlay. The image is oversegmented, meaning there are many small masks rather than one mask for each pear.

L = watershed(Icomp);
overlay = labeloverlay(I,L);
imshow(overlay)

Suppress the shallow minima by applying the H-minima transform. The value for h has been determined by using trial and error. Change the value to see how the h value affects the segmentation result.

h = 30;
Ifilt = imhmin(Icomp,h);

Display the filtered image as a 3-D surface.

surf(Ifilt,EdgeColor="none")
colormap(gray)

Segment the filtered image and display the result. The image contains approximately one mask for each pear in the foreground.

Lfilt = watershed(Ifilt);
overlayfilt = labeloverlay(I,Lfilt);
imshow(overlayfilt)