Effective Content-Based Music Retrieval with Pattern-Based Relevance Feedback

Abstract

To retrieve the preferred music piece from a music database, contentbased music retrieval has been studied for several years. However, it is not easy to retrieve the desired music pieces within only one query process. It motivates us to propose a novel query refinement technique called PBRF (Pattern-based Relevance Feedback) to predict the user’s preference on music via a series of feedbacks, which combines three kinds of query refinement strategies, namely QPM (Query Point Movement), QR (Query Reweighting) and QEX (Query Expansion). In this work, each music piece is transformed into a pattern string, and the related discriminability and representability of each pattern can be calculated then. According to the information of discriminability and representability calculated, the user’s preference on music can be retrieved by matching patterns of music pieces in the music database with those of a query music piece. In addition, with considering the local-optimal problem, extensive and intensive search methods based on user’s feedbacks are proposed to approximate the successful search. Through the integration of QPM, QR, QEX and switch-based search strategies, the user’s intention can be captured more effectively. The experimental results reveal that our proposed approach performs better than existing methods in terms of effectiveness.