I’m kinda new to the Machine Learning world (I only did some basic models during my master) and I would like to receive some suggestions about some good approaches that I could take for my supervised learning task. My data is structured as following: the input is a 3-channels image and my output is a one channel image. Each channel in the input represents a snapshot of a certain quantity in a physical system. Each pixel in the image can be thought as a site in a lattice. The output contains the probability that a given pixel is active. The output image comes from a physical evolution of the input state. What are some good architectures for this kind of problem? Do you have any paper\article treating this kind of problem? Thanks in advance.
If I understand your description, you wish to perform binary (that is to say
two-class) semantic segmentation.
U-Net is a well-established semantic-segmentation architecture.
torchvision includes some pre-built semantic-segmentation
models (but not U-Net), such as fcn_resnet50.
Thanks for the tip, this might be the way to go!