I have a database of defects on plastic films. Defects are burns, holes, and similar things. Some defects are direction specific. A vertical sign on the material represents a scratch done on the material during the extrusion process. Data augmentation is a useful tool, but can be harmful for these defects.
I was thinking about the possibility of implement a custom data augmentation related to the label of the training set. For example: a hole can be rotated without problems, but a scratch not. So I tell to the preprocessing layers how to do the augmentation by passing also the labels.
Implementing this solution gives me a big doubt: this label-bounded data augmentation can cause a leakage of information?