Improving Model Test Accuracy further when it is flat

Hello,

I have a classification model and have used a mini version of ResNext50. It has ReduceLROnPlateau also included.

Model accuracy in test data is stuck with 79% for last 15 epoch although LR getting reduced. My dataset is not skewed and my dataset has variable shape. So the dataloader has resizing to 128. Model has bit over fitting also. Any guidance please how to get out of this plateau?

class1 => height min : 50 width min : 59
class1 => height max : 500 width max : 500
class2 => height min : 87 width min : 78
class2 => height max : 1023 width max : 1024
class3 => height min : 32 width min : 55
class3 => height max : 500 width max : 500

Epoch: 69
Train Loss: 0.320908 Acc: 0.8598
Elapsed 1005.24s, 14.36 s/epoch, 0.38 s/batch, ets 0.00s