Compressing deep quaternion neural networks with targeted regularization