0
$\begingroup$

I am working on semantic segmentation for satellite images using keras and python. It is my understanding that popular models like U-Net require mask images (labels). Are there any unsupervised deep learning models for semantic segmentation that work without mask images? Can we implement unsupervised learning by deep learning models?

$\endgroup$
0
$\begingroup$

Note aside: Unsupervised semantic segmentation is a bit of an oxymoron: semantic segmentation means assigning pixels of an image to labels having particular, semantic meaning, such as "this is a car", "this is a tree", etc. On the other hand, unsupervised means you don't provide any labels to the model. You can do unsupervised segmentation, but it can hardly be semantic if you don't tell the model what kind of semantic information you are looking for.


That said, deep learning provides machinery for performing unsupervised learning: autoencoders. Particularly, recently emerging variational autoencoders with gaussian mixture latent space are a very exciting avenue, and they can surely be used for unsupervised image segmentation.

The article A Survey of Clustering With Deep Learning: From the Perspective of Network Architecture (Min et al., 2018) would be a good place to start looking.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.