Image Segmentation using Mean Shift explained

Sharpie picture Sharpie · Jan 28, 2011 · Viewed 75.1k times · Source

Could anyone please help me understand how Mean Shift segmentation actually works?

Here is a 8x8 matrix that I just made up

  103  103  103  103  103  103  106  104   
  103  147  147  153  147  156  153  104   
  107  153  153  153  153  153  153  107   
  103  153  147  96   98   153  153  104   
  107  156  153  97   96   147  153  107   
  103  153  153  147  156  153  153  101   
  103  156  153  147  147  153  153  104   
  103  103  107  104  103  106  103  107

Using the matrix above is it possible to explain how Mean Shift segmentation would separate the 3 different levels of numbers?

Answer

Dr. belisarius picture Dr. belisarius · Jan 29, 2011

The basics first:

The Mean Shift segmentation is a local homogenization technique that is very useful for damping shading or tonality differences in localized objects. An example is better than many words:

enter image description here

Action:replaces each pixel with the mean of the pixels in a range-r neighborhood and whose value is within a distance d.

The Mean Shift takes usually 3 inputs:

  1. A distance function for measuring distances between pixels. Usually the Euclidean distance, but any other well defined distance function could be used. The Manhattan Distance is another useful choice sometimes.
  2. A radius. All pixels within this radius (measured according the above distance) will be accounted for the calculation.
  3. A value difference. From all pixels inside radius r, we will take only those whose values are within this difference for calculating the mean

Please note that the algorithm is not well defined at the borders, so different implementations will give you different results there.

I'll NOT discuss the gory mathematical details here, as they are impossible to show without proper mathematical notation, not available in StackOverflow, and also because they can be found from good sources elsewhere.

Let's look at the center of your matrix:

153  153  153  153 
147  96   98   153 
153  97   96   147   
153  153  147  156  

With reasonable choices for radius and distance, the four center pixels will get the value of 97 (their mean) and will be different form the adjacent pixels.

Let's calculate it in Mathematica. Instead of showing the actual numbers, we will display a color coding, so it's easier to understand what is happening:

The color coding for your matrix is:

enter image description here

Then we take a reasonable Mean Shift:

MeanShiftFilter[a, 3, 3]

And we get:

enter image description here

Where all center elements are equal (to 97, BTW).

You may iterate several times with Mean Shift, trying to get a more homogeneous coloring. After a few iterations, you arrive at a stable non-isotropic configuration:

enter image description here

At this time, it should be clear that you can't select how many "colors" you get after applying Mean Shift. So, let's show how to do it, because that is the second part of your question.

What you need to be able to set the number of output clusters in advance is something like Kmeans clustering.

It runs this way for your matrix:

b = ClusteringComponents[a, 3]

{{1, 1, 1, 1, 1, 1, 1, 1}, 
 {1, 2, 2, 3, 2, 3, 3, 1}, 
 {1, 3, 3, 3, 3, 3, 3, 1}, 
 {1, 3, 2, 1, 1, 3, 3, 1}, 
 {1, 3, 3, 1, 1, 2, 3, 1}, 
 {1, 3, 3, 2, 3, 3, 3, 1}, 
 {1, 3, 3, 2, 2, 3, 3, 1}, 
 {1, 1, 1, 1, 1, 1, 1, 1}}  

Or:

enter image description here

Which is very similar to our previous result, but as you can see, now we have only three output levels.

HTH!