FCN with spatial encoding

I made images with spots inside of each. For each image, there is a specific distance.
I want to classify my images based on these distances ( I have 10 classes = 10 distances, and thousands of images).

I was thinking of a CNN with some spatial encoding that would then connect to an FCN with the distance function as a loss function. But i feel like it’s not that good of an idea. Are there perhaps some models that can already do this, or perhaps some clustering methods?

Hi just a suggestion if only the relative spots are relevant you can use graph based approaches like graphconv networks from pytorch geometric

1 Like

Just to see if I understood what you meant correctly. I would have to first detect the spots positions and feed that to the graphconv and classify based on that?

yes build a graph with spot positions as nodes and feed the formed graph as input you can even use the relative distances as weight of edges. use this graph as input.

And for getting the spots you can use opencv based contour detection approaches