What neural network architecture to choose to recognize objects by mask?

Question

Python

I have a large dataset, such images (5000x5000px) and masks for them:

Is there any architecture or ready-made implementation of a neural network that could generate such a mask yourself in the future. The catch is that the standard implementation of R-CNN only recognizes rectangular regions defined through two points (I'll be wrong).
The closest thing for me is Python->Keras, but other tools will do.
UPD: The answer below solves the problem perfectly. As an alternative, I look towards Pix2Pix .

Answer 1

U-net , there are tons of ready-made implementations.
Train on random crops, predict by window.