Suppose I have a set of characters with both natural scene and synthetic images, and another set with only synthetic images. I'd like to make a classifier which is trained on only on this data and tested on natural scene images of characters from the second set. Does anyone know of an approach that give good test accuracy in such case?
Example:
Synthetic digits:
Natural scene digits:
Synthetic thai characters:
Natural scene thai characters: (not present during training)
