Episode 6, Exercise 3 labelling with connected components

[wangliyuan72]

Brief description

Connected component labelling is a prevalent way to label pixels that could be identified as the same object. To perform such analysis, it is important to define 'connectivity' which pixels are considered direct neighbours of a pixel. I think neubias's page on connected component is great to use. It demonstrated 1 and 2 connectivity in both 2D and 3D.

From prior exercises, students now have binary images from different thresholding approaches. Briefly, with binary images as input, we use connected components analysis by defining connectivity, producing a label image with separated objects. The exercises use label function from skimage.measure. skimage.measure.label(label_image, background=None, return_num=False, connectivity=None)

Input:

• binary images (typically background and foreground regions)
• We use a small sub-region of a binary image from previous step. (side note: EPFL course demo-ed the concept with a 6-by-6 made up image.
• Students will be able to say how many objects in that region.

2nd step:

• we introduce the concept of connectivity as we need to define which pixels are considered direct neighbours of a pixel. Essentially, do we consider the diagonal pixels as part of the objection x (demo-ed below).
  labels_2D_conn1_image = label(binary_2D_subregion, connectivity=1)
labels_2D_conn2_image = label(binary_2D_subregion, connectivity=2)
• we can use matplotlib.pyplot.imshowto visualise them.
• Probably worth mentioning that the label image contain objects (each has its own assigned integer).

1-connectivity     2-connectivity    
     [ ]           [ ]  [ ]  [ ]            
      |               \  |  /               
[ ]--[x]--[ ]      [ ]--[x]--[ ]      
      |               /  |  \            
     [ ]           [ ]  [ ]  [ ]

We can mention the following equivalent nomenclatures

• 2D: 1 connectivity = 4 connectivity
• 2D: 2 connectivity = 8 connectivity
• 3D: 1 connectivity = 6 connectivity
• 3D: 2 connectivity = 26 connectivity

3rd step:

• to visualise label images by colour coding and overlaying raw data. [skimage.color.label2rgb](Probably worth mentioning that the label image contain objects (each has its own assigned integer).)
  image_label_overlay = label2rgb(labels_2D_conn1_image, image=raw_image, bg_label=0, kind='overlay')

4th step:

• we can count the number of objects in the label image by len() of the resulting array of `numpy.unique(labels_2D_conn1_image)

Learning objective(s)

Covers Objective Labelling with Connected Components.

I think the first two objectives can be achieved by working on (segmented) binary images (2D or 3D) to identify objects by defining connectivity=1 or 2. Less connectivity, one will have more separation; more connectivity will result in more merging of objects.

with label2rgb function, Learners will be able to overlay colour coded labels over desired image (we can define this to be the original image).

Volunteer(s)

@terezbelinova @marcodallavecchia (Hi again, feel free to add any comments). Thanks
@wangliyuan72

[marcodallavecchia]

Thanks for the contribution @wangliyuan72!

I think both the NEUBIAS link and the main ideas of what you are proposing are very relevant for this learning objective. I only feel the exercise proposal itself is a little underdeveloped.

Do you think you would be up for defining a few more details of how the exercises would go?

For example we could rephrase it into a clear ordered or bullet point list which narrates the main points to be covered by the exercise from binary to final labeled objects (with the corresponding data and functions to be used). You can always either Edit your own issue description or simply reply to this Issue with a comment and I can offer to integrate the final proposal in the description myself.

[wangliyuan72]

Hi @marcodallavecchia, I updated my original post, please take a look and let me know what you think. Cheers