To avoid conflict between algorithmic efficiency and recognition effectiveness in face recognition,this paper proposes a zero-norm sparse coding algorithm. The algorithm uses zero-norm to describe sparsity of a sparse coding model and applies a strategy of continuous extension of discontinuity points to speed convergence. A test based on the ORL database show that the algorithm is more efficient in adjusting sparsity so that the computation time is reduced, and gives higher recognition rate as compared with the methods of nonnegative sparse coding and non-negative matrix factorization with sparseness constraints.