Research On Content-Based Music Information Retrieval With Relevance Feedback

Digital music is one of the most important multimedia types. The popularity of digital music is rapidly increasing thanks to improved digital music processing technologies and convenient availability facilitated by the Internet. This gives rise to a need for making an efficient scheme to retrieve the music that the user is interested in. At present, applications that manage music data like Google, Baidu usually utilize textual meta-data, so that computers can understand. The meta-data often contains meaningful descriptions of the music, such as title, artist, genre, emotion and so on. While manually annotating these music data, which has been using is time-consuming, costly and subjective. In recent years, content-based music information retrieval has received widespread research interest. However, due to the existence of the so-called Semantic Gap, the performance is not satisfactory. Motivated by this, relevance feedback is introduced to the area of music information retrieval. This dissertation is dedicated to the study of the key techniques of content-based music retrieval with relevance feedback.We first give an introduction the state of the art of feature extraction in music content understanding. As for music content representation, we discussed the extraction of some features used in speech recognition and some features only used in music content representation. A novel method for beat detection based on Const Q Transform features and dynamic programming is developed. This part is important foundation of music retrieval.Based on the feature extraction, support vector machine based music retrieval with relevance feedback is focused on. First, to address the problem that one-class support vector machine can only learn from the positive samples, a novel approach based on one-class support vector machine and relevance feedback is presented. Through exploiting the information from positive and negative samples provided by the user, better performance can be achieved. Second, we present a novel multi-sample selection strategy for two class support vector machine based relvance feedback. The key of the strategy is how to reduce the redundancy among the selected samples such that each sample provides unique information for model updating. To this end, we use the distance diversity and the density as the measurement of the distribution of the selected sample set and choose the set of music samples to effectively maximize the distribution. Experimental results with two different music databases demonstrate the effectiveness of the proposed strategy.With the application to the music annotation, in order to reduce the human effort as much as possible during annotation, a method based on multi-class support vector machine is presented. To provide multiple samples to the user for annotation and learn multiple models simultaneously, a multi-sample selection strategy with active learning is proposed. Compared to the conventional techinques, our method can significantly improve the performance of the annotation.