What is an effective image data padding technique?

I am trying to classify figures in Japanese paintings, and I am using efficientnetb0, but the accuracy is not improving.
I am using HorizontalFlip and RandomRotate90. Are there any other effective means?