Cross-modal Music Retrieval Based On Canonical Correlation

With the rapid development of Internet and multimedia technology,a large number of digital music has been expanded on the Internet,allowing users to have a lot of options. Text-based retrieval is the traditional technology in the search engine, which depends on the musical attributes only to indirectly receive music. It is clearly that the music retrieval algorithm like this has been unable to meet the growing intelligent, personalized search needs, So content-based music retrieval technology came into being.A query by humming algorithm is the mainstream way in the study of content-based music retrieval technique, but this approach is so specialized that it is difficult to widely spread to the general users.With the development of the Web2.0 technology, semantic-based music retrieval methods are gradually stepping onto stage. On some music sites such as Last.fm, users can apply music labels which usually have relatively high semantic information to mark or describe their interested songs. However, relying on labels only to retrieve may overlook the similarity in the content level between different multimedia data. So, the question is that how can we take features from different modals into consideration to complement each other between the various modes of information to improve the retrieval accuracy, a novel search method is required to handle data from different modes.This paper aims to study this multimedia data retrieval methods, namely, cross-media retrieval.Cross-media retrieval technology is inseparable from the content-based multimedia retrieval techniques. To make up for the shortcomings of a single modal retrieval and improve the efficiency of information retrieval, different types of multimedia data is needed to carry out a comprehensive analysis. Existing music retrieval methods are mostly designed for data of single modality, this paper aims to provide a cross-modal music retrieval mechanism based on canonical correlation.First statistical method is applied to analyze the underlying relationships between the text features and music content features,then subspace mapping is used to solve the heterogeneous problem between different feature vectors,at last Euclidean distance can be employed to measure the cross-media similarity, in order to further improve the retrieval accuracy,weight allocation is used to optimize search results.The experimental results show that the proposed method in this paper can get a better result in music retrieval.