Pattern Segmentation using U-Net - Validation loss increasing

Hello All, I am trying to do a segmentation for a pattern recognition. I have only 2 class label ( Masked Image)
Pixel Value - 0. , Class Index - 0 ( Pattern Area , Black Region as shown below)
Pixel Value -1. , Class Index -1 ( Back ground Area)

I have tried training with different parameters , below are my observations:

  • The model over-fits. The Pixel prediction accuracy in the 1st epoch is high. As the epoch increases, the validation loss increases & pixel prediction accuracy decreases. While my training loss decreases continuously

  • Loss Function → CrossEntropy , Model → U Net

  • Training Images ~ 2600 images & ~600 validation Images

  • I have tried to train with small image size(64x64), the training trend does not change

  • I have tried to give a weight decay of 2 , I get only one class prediction ie., the back ground class pixel prediction all over the image

Masked Image looks something as shown below. Sample

I have trained the model for Max of 10 Epochs

Kindly let me know what would be the likely issue for this. Also any suggestions to get the validation error decreasing.

Note : I want to use AI for this, not any openCV techniques

Thanks in advance