Datenbestand vom 13. Juni 2019
Tel: 089 / 66060798
Mo - Fr, 9 - 12 Uhr
Fax: 089 / 66060799
DER VERLAG IST IN DER ZEIT VOM 12.06.2019 BIS 23.06.2019 AUSCHLIESSLICH PER EMAIL ERREICHBAR.
aktualisiert am 13. Juni 2019
978-3-8439-0614-2, Reihe Informatik
Extraction and Evaluation of Mid-level Features for Semantic Music Analysis
211 Seiten, Dissertation Universität Passau (2011), Softcover, A5
Many computer users nowadays have access to huge music archives and they have the possibility to listen to extensive audio collections also when they are mobile. However, the search capabilities within these collections are still rare. The automatic computation of such criteria directly from the music is covered by the scientific field of music information retrieval (MIR). A large number of techniques in MIR use feature extraction and classification methods. Feature extraction deals with the computation of matrices from musical pieces, which represent the musical properties. The classification technique uses the extracted features and categorizes them into classes based on previously trained feature examples. Two types of musical features have existed so far: low-level and mid-level features. Low-level features are based on simple mathematical rules and they were comprehensively described in literature. The extraction of mid-level features is much more complex and semantic knowledge is added in order to adapt them to a certain musical property. This thesis examines their state of the art of three different directions - rhythm, dynamics, and harmony. The advantages and disadvantages of these approaches are illustrated in this thesis. Additional improvements are proposed, which lead to a higher classification accuracy. Furthermore, the most promising approaches of the three directions are used for genre recognition. This thesis examines the difference between low-level features and mid-level features, shows the advantages of the mid-level feature usage and possesses its barriers.