Design Of The Content-Based Polyphony Music Retrieval System

How to find the required data from the growing multimedia resource accurately had become a research hotspot in the field of multimedia retrieval. The research of related technology will promote the development of Human-Computer Interaction. Compared with the Content-Based Image or Video Retrieval, Content-Based Audio Retrieval (CBAR) developed more backwardly. Query by Humming/Singing (QBHS), which is a branch of Content-Based Audio Retrieval, becomes one of the most natural and convenient music retrieval styles. When a user hums a piece of melody, music retrieval system can search the similar song from music database. Most existing music retrieval systems have been developed in the basis of MIDI files, instead of polyphonic music such as MP3or WAV files.On the basis of researching the key technologies and difficulties about the polyphonic music retrieval, the thesis completed the humming input endpoint detection and target music division, improved the existed pitch extraction algorithm based on fundamental frequency matrix, induced a dynamic time warping (DTW) algorithm fitting to music retrieval, realized the extraction and approximate matching of musical feature. At last, the thesis designed and implemented a polyphonic music retrieval system named Music-Retrieval.The main research contents of this thesis are as follows:(1) On the basis of probing the theory of music retrieval and the knowledge of music theory, this thesis chose the melody (pitch) that used in music retrieval widely as the expression of musical feature, designed the overall framework of the polyphonic music retrieval, as well as summarized the difficulties of the polyphonic music retrieval and the focus of this paper research.(2) This thesis discussed the knowledge of musical signal preprocessing, realized endpoint detection of the humming inputs, which choosing the algorithm based on energy-spectrum, and completed the segmentation of target music, which can decrease a large number of workload in the late stage of musical feature extraction and matching.(3) This thesis described the faults of some pitch extraction algorithms based on time or frequency domain. On the basis of studying the existed pitch extraction algorithm based on fundamental frequency matrix, this paper improved and perfected this algorithm, which can extract pitch form target music when having no ideal to construct the fundamental frequency matrix, and reduced the complexity of calculation.(4) This thesis analyzed and discussed the merits and drawbacks of DTW algorithm that used in voice recognition widely, induced a DTW algorithm fitting to music retrieval, completed musical features matching. The new algorithm can correct the tone errors of the humming inputs and decrease the time of calculation.(5) This thesis completed the functional design of the Music-Retrieval system, programmed and realized this system using VC++. At last, this paper collected a few humming inputs as experiment data, and completed experimental evaluation about Music-Retrieval system, including the validity of the algorithm, the performance, speed and accuracy of the system, etc.