U-Net Segmentation Blobby


I am using U-Net Image Segmentation to detect buildings in satellite images (label 0: no building, label 1: building). If I use the pixel-wise error loss, the predictions are very blobby and the borders are not very accurate. Is there a way to adapt the loss function to make the borders more accurate and the predictions less blobby?

Can someone help me?

Sorry, but nobody from the fast.ai community seems to care about pixel-wise weighted loss functions