Nowadays, dictionary learning has become an important tool in many classification tasks, especially for images. The tailor-made atoms in a dictionary are trained for the reconstruction of the test sample. In the classification, atoms are associated with different classes from several subspaces such that the test sample is labeled according to the distances of each subspace. However, it is hard to fix the number of atoms to obtain the optimal result for each scenario since the optimal subspaces required are different. To improve the classification performance as well as the robustness, we proposed subspace-level dictionary fusion (SLDF) to construct a dictionary-based classifier. A full-size dictionary and a locality-constrained dictionary are constructed in parallel. Then, the reconstruction coefficients of the two dictionaries are obtained, which leads to a pair of distances between the test sample and the subspaces. Finally, a decision is made according to the pair-wise fusion of the distances. The experimental results on multimedia datasets from distinct categories such as image, text, and audio show that the proposed method outperforms other state-of-the-art dictionary-based classification methods with accuracies of 99.74% (image), 83.96% (Text), and 87.07% (Audio).