Keras: how can I implement the fcn structure?

Above image is the FCN structure. I want to implement it by keras. There is one input.The Upsample block genrate 4 different resolution predicted image,so it is needed to provide corresponding 4 different scale ground-truths, and the final loss is:

My question is how can I implement this loss with keras?
Is this a problem about multi-output model?

Thanks very much!