Research On The Content-Based Piano Music Retrieval

With the rapid development of the network technology and the multimedia technology, people are able to contact conveniently more and more music. The problem is how to find the music quickly that is people's need in the vast music. Therefore, the study of music retrieval becomes very important and it is concerned gradually by people. The traditional music retrieval aims at monophonic music, such as humming, but there is a little research on polyphonic music, such as piano music. However, with the improvement of people's living standards, more and more people, especially children who have strong interested in learning piano, and the piano has the most extensive range of pitch in all of the instruments. So, the study of the content-based piano music retrieval is important to people's entertainment life and research on polyphonic music retrieval.Research on the content-based piano music retrieval includes the establishment of audio features database, feature extraction, retrieval matching and similarity calculation. This thesis has the following studies:1. Establishment of audio features database. Standard database of music pitch was created by reading the MIDI format music and the N-grams index algorithm was introduced to establish characteristic index database taking the treble as the key word.2. After analyzing the adaptive threshold on the peak extraction of onset detection, the instability of threshold resulted from the median filtering was found. By analyzing the characteristic of average filtering threshold, the method of combining the median filter and the mean filtering was researched to improve the stability of adaptive threshold. The improved algorithm and original algorithm were simulated; the results show that the improved algorithm could receive better effect than original algorithm.3. For existing multiple fundamental frequency estimation algorithms can not extract pitch of piano music accurately, a method of extraction the sequence of candidates for pitch was presented to instead the accurate pitch based on the algorithm of harmonic peaks. The sort of each pitch in the sequence of candidates for pitch according to each pitch's magnitude of frequency domain, in accordance with the different location of candidates for each pitch in sequence, different locations of the pitch was given different scores. Algorithm of similarity calculation was presented based on these scores.4. According to the research on the algorithms of piano music retrieval, a content-based piano music retrieval demo system was designed and realized. The experimental results were analyzed about the search time and retrieval accuracy in the case of the different length of input music and some suggestions were given to users about the length of input music in order to get higher retrieval accuracy.