Audio Fingerprinting Retrieval Systems Based On Compressed Suffix Array

Alone with the rapid development of Internet, the popularization of electronic devices and theemerge of High-capacity memorise, people can get more and more information with an astonishingspeed. Especially in the last few years, the mobile Internet technology grows fast and the mobileterminal becomes more and more popular, people can upload their information and download theirneed for services or information at any time. The audio content take a significant part of theseinformation and services. How to search and manage these massive music data has become a newtopic. However, the management of massive music data is a very tedious and error prone task. Inrecent years, CBMR (Content Based Music Retrieval) has become an important topic in processmassive multimedia data under network environment. Together with image retrieval and videoretrieval, it has become the hot spot in the research of content-based multimedia retrieval.As a kind of message digest (one-way hash function), audio fingerprinting converts an audiosignal into a relatively compact representation by using acoustical and perceptual characteristics ofthe audio signals. It can be used in copyright protection, audio content identification, contentintegrity verification and some other fields, has a very important significance.This thesis introduces the basic concepts and background about music retrieval, also describesthe frame of the retrieval system and the related technology. Focusing on the study of the indexcompression algorithm in music fingerprint database and two cores in retrieval system, featureextraction and retrieval algorithm. Finally realized the system based on those research and thecomparative test was designed to verify the effectiveness of the system. The main research worksare as follows:(1) In view of the current massive music data and it produce a large amount of storage spacefor index of fingerprints, we propose using compressed suffix array to compress the index to solvethis problem. Taking advantage of the fact that the repetitive characters occur frequently in higherbits of the sorted audio fingerprint data, the proposed method compresses the index by encoding the8-bit data sequences by Run Length Encoding. Vertical Code is also used to compress the array.(2) Proposed a method to retrieve the MP3lossy compression format music by using MFCCfeatures.(3) Using Kullback-Leibler Divergence and Earth Mover’s Distance (EMD) to compute musicsimilarity. And finally we proved that in the same conditions the precision of EMD was higher thanthat of KL divergence by designing the contrast experiment.(4) According to the fingerprint extraction and retrieval algorithm researched in the paper, amusic retrieval system prototype is designed and implemented. With contrast to the traditionalsystem, we proved the effectiveness of the new system.