Abstract: The existing binary feature learning algorithms for face recognition don’t distinguish the original spatial features. In order to solve the problem, this paper introduces the sparse constraint based on L₂,₁ norm to binary feature learning algorithms. We used the L₂,₁ constraint to induce the generation of sparse projection matrix in iteration, which improved the contribution of important features and reduced the influence of secondary features. Considering the computational cost, the training set was decentralized to replace the bit balance term of the objective function to simplify the calculation, and the rationality proof and iterative formulas of the objective function were given. Experimental results show that, compared with other similar algorithms, the proposed algorithm achieves better results on three public face databases: FERET, CAS-PEAL-R1 and LFW.