You can use kmeans (BGR data) with 4 clusters :
with false color in left image yellow color is only shadow. Yellow is only for shadow. Right image is connected components images
source code https://github.com/LaurentBerger/Trai...
OK can you post some snippets or better a complete code program so I can see ?
the kmeans is done on what data : luminosity, saturation, hue ?... and look at the tooth upside there is a part of shadow which has been tagged leaflet
I coded kmeans with 4,5 and more clusters but it didn't work because parts of leaflet have been tagged background and parts of background (shadows) have been tagged leaflet, this is not the solution, another idea ? I think the solution is more complex like work with textures or nearest canny edges
the method with saturation doesn't works : leaflets parts are tagged background (which was not the case with luminosity) and some parts of background are tagged leaflet, it is worth than luminosity
Asked: 2019-02-20 11:32:27 -0600
Seen: 1,174 times
Last updated: Feb 24 '19
I like to do some grabcut after getting the ROI from Otsu but actually I think to ride out of the shadow is a really difficult task.
A person on channel #opencv advised me to use saturation for grayscale transformation before Otsu binarization instead of luminosity , what do you think about this ? The kmeans seems not working because how to segment after ?
You can simply use a mask to segment image using labels images. with 4 labels you can extract :
mask =labels==1 connectComponnent(mask,...
if I tag the parts it is that I already know the parts but what I want to do is to discover them automatically