Cascade Classification And Multi-label Classification Used In Chinese Folk Instruments Recognition

Currently, the capability of computer to recognize music events is severelylimited when compared to human ability. The recent development of digital audioand network technologies has enabled us to handle a tremendous number of musicalpieces and therefore efficient music information retrieval (MIR) is required. Musicalinstrument recognition will serve as one of the key technologies for sophisticatedMIR. Despite the importance of musical instrument recognition, studies have untilrecently mainly dealt with monophonic sounds. In this thesis, we address recognitionof musical instruments in polyphonic music. At the first stage, an efficient algorithmis proposed for selecting the most appropriate signal features for a givenclassification task. The method is based on MPEG-7standards in acoustic features,adding some features modified with time property. Comparing the performances oftwo created classifier models——with/without adding the new characteristic value,the latter is better. Our system for the classification of audio signals according toaudio category is a tree-like structure. The flat classification approach, typicallypredicting classes only at the leaf nodes, is commonly used. However, this approachdoes not explore information about parent-child class relationships present in theclass hierarchy. We employ top-down approach for the pre-built hierarchicalclassification that takes into account by using local information named local classifierper level. The experiments performed on Chinese national musical instrument dataset and the result show that hierarchical classification achieves high classificationaccuracy efficiently compared of the flat classification approach. Finally, the manualof catalog in musical instrument is modeled as a multi-label classification task, wherea piece of musical instrument may belong to more than one class. Two algorithms areevaluated and compared in this task. Experiments are conducted on a set of50256frames with3clusters of music catalog based on different playing methods. Resultsprovide interesting insights into the quality of the discussed algorithms.