Multimedia Processing Techniques for Retrieving, Extracting, and Accessing Musical Content ; Techniken der Multimediaverarbeitung zur Suche, Extraktion und den Zugriff auf musikalische Inhalte
You can manage bookmarks using lists, please log in to your user account for this.
Media type:
E-Book;
Electronic Thesis;
Doctoral Thesis
Title:
Multimedia Processing Techniques for Retrieving, Extracting, and Accessing Musical Content ; Techniken der Multimediaverarbeitung zur Suche, Extraktion und den Zugriff auf musikalische Inhalte
Contributor:
Balke, Stefan
[Author]
Published:
OPUS FAU - Online publication system of Friedrich-Alexander-Universität Erlangen-Nürnberg, 2018
Footnote:
Diese Datenquelle enthält auch Bestandsnachweise, die nicht zu einem Volltext führen.
Description:
Music constitutes a challenging multimedia scenario. Besides music recordings, there exist a number of other media objects including symbolic music representations, video recordings, scanned sheet music, or textual metadata. Developing tools that allow users to retrieve information from different types of music-related data is central to the research area known as Music Information Retrieval (MIR). This requires techniques from various engineering fields such as digital signal processing, image processing, data management, and machine learning. In this thesis, we develop novel multimedia processing techniques and explore their capabilities and limitations within different complex music scenarios. The thesis consists of three main parts. In the first part, we consider retrieval scenarios within a Western classical music setting. For example, given a short monophonic melodic theme in symbolic notation as a query, retrieve all corresponding documents in a collection of polyphonic music recordings. In another related retrieval scenario, we aim to link the score of musical themes, scanned from book pages, to their symbolic counterparts given in MIDI format. Both scenarios require mid-level feature representations derived from different media types, as well as robust retrieval techniques that can handle extraction errors and variations in the data. The second part of this thesis deals with the extraction of musical parameters such as fundamental frequencies or musical pitches from audio recordings. In this context, a general goal is to reduce the variations in the degree of polyphony between monophonic queries and polyphonic music databases. In our computational approach, we propose a data-driven method based on Deep Neural Networks (DNNs) which aims at enhancing salient parts from jazz music recordings. As an example application, we employ the learned model in a retrieval scenario where we take a jazz solo transcription as a query to identify the corresponding music recording. In the third part, we explore the ...