Creating custom loss function for regression in CNN

Question

I am trying to make my CNN regression network run smoothly, but output is far from ground truth. The input is a binary image (600x600) which the background is b