Retrieval And Recommendation System Of Resources Based On Deep Belief Networks

Audio, video and other multimedia data has become the major information carriers in the field of digital information processing. It is increasingly important to process and analyze of them effectively and rapidly. Music mood classification of this study is a research hotspot of audio classification. It’s essentially a pattern recognition problem. The main work and innovation of this paper are summarized as follows:Firstly, this paper introduces the background of music emotion recognition and also the research and development of deep learning. Then, this paper describes in detail the structure and learning process of the restricted Boltzmann machine (RBM) and a deep belief network (DBN) model. This paper proposed a model based on DBN for the fusion of the multi-modal features, audio and lyrics of music. Thus the learning algorithm must be improved for the fusion. Considering that the bottom-up layer wise learning of DBN is based upon conditional probability activation mechanism, multi-modal features can be learned by associated activation probability. We make it in by the layer wise learning of the correlation between the multimodal the hidden layer mapping function. In this way, the fusion of the hidden layers of each mode not only preserves the distribution of the pattern features, but also provides the convenience for the feature fusion.It is the first to use DBN to learn multi-modal features. The fusion of the multi-modal features solves the problem of feature reconstruction problem in music mood classification, and improves the prediction accuracy and stability of music emotion classification system. The multi-modal learning based on DBN is realized by stacking restricted Boltzmann machine with some improvements of original learning algorithm, adding the association based on distance measure with experiments to improve it.