Predominant value in data label: how to assign equal importance for every label despite their amount

[Marco379]

Hello, 

I am training a CNN for pixel-wise classification. My ground truths pixels can assume 4 values, 0 OR 1 OR 2 OR 3. 

In my dataset labels, the value 0 is predominant with respect to the others. 

Do you know a way to impose a different importance or weight to a certain ground truth value during training?

Because the accuracy remains very low (about 0.25), and I see that as the training proceeds, the net layers parameters are updated toward "zeros", that in my opinions represents the nearest local minimum (all zeros yield an output map of all zeros that is not-so-terribly-accurate since the predominance of 0's in the ground truth).

Many thanks,

Marco