Connected component analysis

[tischi]

@manerotoni @ssgpers @haesleinhuepf @dlegland @stelfrich

I am working on teaching material for connected component analysis.

Just recently I realised that the input to a CCA does not need to be a binary image, but could be multi valued.
The algorithm will just assign the same label index to neighbouring pixels of the same value.

I tested this in MLJ and also there it works for multi valued images even though the menu entry in Fiji suggests that the input should be a binary image.

Is that a consensus that CCA should accept also multi valued images as an input?

[dlegland]

Hi,

IMHO, the basis of CCA is to transform a binary image into a label image. So, if there is a consensus, it would be more to restrict the definition to binary images.

It may be possible to extend the definition to multi-valued images such as label maps, but this complicates the definition and the implementation. The scikit-image library allows applying CCA to label maps for example.

Also, I think that if it works in MLJ, it does not handle the case of contiguous regions, so handle with care!

[tischi]

Hi @dlegland,

IMHO, the basis of CCA is to transform a binary image into a label image.

Some weeks ago I would have agreed, but at I2K @frauzufall presented a workflow for cell segmentation ground-truth annotation in LabKit, which relied on the fact that downstream CCA would also work on multi-valued images. The idea was that one could draw the same label index for all non-touching cells and only needed distinct labels for touching cells. This is much more convenient than having to draw a different label for each and every cell.