U-Net Background Subtraction Exploration

[kevin-seeger]

I'd like to take a look at using a U-Net to remove the background from images prior to feeding them into some other classifier. The idea here is that we have a dataset of empty images that we can utilize, and training a model to leverage this could be helpful.

The U-Net architecture is popular in biomedical fields where the background tends to fairly uniform like in histology slide images or MRI images, for example. I have some doubts as to how well this will translate to backgrounds that aren't completely static, but I think this is worth exploring.

I have found this article that I'm using as a reference, and will update this thread with other resources as I come across them. I know this wasn't explicitly one of the techniques mentioned in our proposal, but I don't think that the proposal holds us to anything specific.

U-Net Article: A Comparative Study of U-Net Topologies for Background Removal in Histopathology Images

Reference for Implementing U-Net architectures: link

[kevin-seeger]

And here is the original U-Net paper. I'm going to try implementing this architecture first.

[kevin-seeger]

Pretty cool, got a U-Net model running tonight, seems to do a decent job generating backgrounds. I'll have to do some more exploration to see how well it will segment the animal from the background, but for now looks like it can do something: