A Research On Key Technologies Of Music Melody Automatic Extraction And QBH System

With the increasing applications of multimedia data on the Internet and multimedia databases, there are more and more researches on automatic classifying, indexing and retrieving multimedia data. This is particularly true for music files, since much more online music storing and searching applications are available and they need an efficient mechanism to organize these massive data. Text-based retrieval is only effective for music files with metadata. Content-based music information retrieval is to access the desired music files by melody, rhythm, timbre and other difficult-to-extract layers of significance.This thesis mainly focuses on the key technologies on content-based music retrieval, mainly involving musical melody feature extraction, musical melody matching, the structure of musical melody database, the whole retrieval system constructing and many other issues. Currently at home and abroad, researchers have made great progresses on these technologies in terms of theory, and developed many relate application systems. However, there are many problems with extensiveness, automation and scalability in these existing systems. Therefore the research work in this thesis is related to these problems. The studies are mainly on automatic construction of music database, melody automatic extraction method from polyphonic MIDI and system's implementation for MIDI.First, through modifying an open source crawler (Larbin), I realize an offline system to automatically construct the music database. It enhances the whole system's automation and the scalability of the music database.Second, the thesis explores the theory of melody similarity and designs a new representation method for melody. I propose and realize an algorithm for extracting the melody from polyphonic MIDI based on the melody similarity computation. Experiment results show that the recall and precision are better, and prove the effectiveness of the algorithm.Finally, the thesis constructs a musical retrieval system based on humming. Experiment results show the system is efficient.