In the dictionary learning algorithms, the model by the multi-vector representation can obtain better classification performance and more robustness than that by the single vector representation. In this study, we use the combined representation fused multiple vector representations and reasonable weighted logarithms sum schemes to improve the performance of the dictionary algorithm. Experiments on public face datasets verify that dictionary learning algorithms applied with proposed method has higher accuracy and robustness. It illustrates that the various potential appearances of observed objects generated by fully mining and utilizing the diversity of representations are beneficial to improve the performance of images classification.