Autoencoders with Intrinsic Dimension Constraints for Learning Low Dimensional Image Representations