Why do deep neural networks generalize so poorly to small image transformations?

Is there a way to generate these variations and feed them into the network as training data?