In the paper we explore the two important parts of music emotion classification:feature selection and classifier. In terms of feature selection, single feature cannot fully present music emotions in the traditional algorithm, which, however, can be solved by the multi-feature fusion put forward in this paper. Specifically, the sound characteristics and prosodic features are combined as a symbol to express music emotion. In the classifier selection, the deep belief networks are adopted to train and classify music emotions, which had a better performance in the area of audio retrieval. The results show that the algorithm performs better than the single feature classification and SVM classification in music emotion classification.