Towards improving discriminative reconstruction via simultaneous dense and sparse coding