Abstract: Sparse representation-based classification model contains non-connected group structure. The proposed face recognition method utilizes non-connected dynamic group sparse to improve representation ability of model. The method employed combinatorial search to obtain all possible base-blocks, including non-connected base-blocks, and constructed dynamic group structure by union of the base-blocks. Meanwhile the method divided high dimensional data into blocks according to human category, and performed combinatorial search within block in order to avoid combinatorial explosion. Furthermore the method adopted coding complexity to measure structural sparsity of coefficients, and analyzed calculation methods of coding complexity for several sparse structures; then achieved sparse reconstruction of non-connected dynamic group structure based on structured greedy algorithms. Finally, the face recognition experiments on the AR, Extended Yale B and CMU-PIE databases demonstrate the effectiveness and stability of the proposed method.