That's a bunch of noisy pictures of very similar numbers. You can add noise to the pictures you're training your CNN with; it's not really going to help. It's still texture without structure or context.

You can show me novel symbols, with me only looking at a single example of each for a few seconds, and I can manage good categorization.