Trading off Image Quality for Robustness is not Necessary with Regularized Deterministic Autoencoders