Validaiton result is not good,how to improve it?

Hi,i want to use Unet segment target in different dataset.In Unet,using GroupNormalization,LeakRelu,maxpooling and conv,traing opt is Adam,loss is mutildice,accu is mean dice.
first i segment five target,the train and validation performence are both good,look first line:
second i segment three target,the train performence is good,but validation is not well,look second line:

can anyone know how to solve it?

thank you!!!

you need to add the regularization i.e L1, L2 . If the results were not improved you can increase the dataset by using data augmentation

i have add data augmentation with zoom, pan, rotate, flip operation,i will try regularization L2 to Adam opt.

you also need to add the dropout layer as your model is clearly over fitted

ok,i also add dropout layer after conv operation.

i add dropout,but the result is also not good.

also increase the learning rate may be it will help

So I don’t know if you already do more, but you seem to mention techniques that would be typically used for classification problems where the image has a “global” label.
One of the ways to look at pixelwise segmentation (e.g. with UNets) is as a pixelwise classification problem. Then, perhaps not as obvious when you don’t explicitly entertain the pixelwise classification view, class balance is important. We briefly discuss this in Part II of our book and a long time ago, there was also this thread:

Best regards