Methodology doesn't support the conclusion.
Try using images that it wasn't trained on.
Try using images that it wasn't trained on.
This seems like a man / women / skinny / fat detector.
If you just classified people into those categories I wonder if you would achieve a similar level of accuracy.