Efficient Retrieval For Massive Audio Based On Content

With the wide popularity of the Internet around the world since the new century, the rapid development of audio codec technology and the birth of high-capacity storage medium, the quantity of digital audio resources in network render index level of growth. Huge amounts of digital audio resources in the network bring greatly convenience to people, however, because of the lack of standardization in the management system of digital audio and the faultiness in the regime of copyright protection, Internet users are free to upload and download digital audio resources or change the content of the audio, this behavior seriously violated the legitimate rights and interests of copyright owners of the digital audio resources in the internet virtually. Copyright protection of digital audio has caused extensive concern of all sectors of the society, and it has become an important problem to be solved urgently.According to demands of the major science and technology project of the general administration of press and publication named "Research and development project of digital copyright protection technology" and the support topic of the national ministry of science and technology called "key support technology research of digital copyright service", this topic mainly studies the relevant key technology of the record and efficient retrieval of audio features, and finally realizes that the average time of retrieving and positioning an unknown audio clip using a typical server with100,000digital audio is less than1second, at the same time the retrieval accuracy need to be ensured above90%. Research of this topic has the promote function and significance that not to be ignored to the normative management of the huge amounts of digital audio resources、effective copyright protection of digital audio and fast and accurate access of digital audio content under the network environment.Firstly, the topic has carried on the detailed elaboration of the domestic and foreign research present situation of the content based audio retrieval system. Through the comprehensive summary and analysis to the existing audio fingerprint extraction method and the related rapid retrieval method, this topic focused on the in-depth discussion of the related fast retrieval method about the classic Philips audio fingerprint. Finally we designed an efficient retrieval system for massive audio based on Philips fingerprints, and a large number of experimental verification has been done. The main contribution of the project are:1) With introducing the bag-of-features algorithm on the basis of the Philips fingerprint, an efficient and robust intermediate fingerprint with the fingerprint data reduced multiply compared to Philips fingerprint has been proposed to make a filtering retrieval, it can filter a large number of irrelevant audio rapidly in a short time;2) A fixed interval sampling matching method with thresholds has been designed accordingly to sharply reduce the similarity matching calculation in the process of the retrieval, and effectively promoted the filtering speed of the middle fingerprint;3) Combining with the Fibonacci hash indexing algorithm、the intermediate filtering fingerprin、Philips fingerprint and the fixed interval sampling matching method with thresholds, an efficient cascade audio filtering retrieval system has been designed and realized.By a large number of repeated experiments, the intermediate filtering fingerprint based on BoF and Philips fingerprint has been proved that it possesses high filtration velocity and amplitude while the recall rate and accuracy of retrieval is ensured; a fixed interval sampling matching method with thresholds can effectively improve the filtering speed and speed of retrieval based on slightly reduce the amplitude of the filter; running on the final designed efficient cascade audio filter retrieval system, the average time of retrieving an unknown10-second-long audio clip using a typical PC with100,000audio is just0.15seconds, but the retrieve recall rate can reach more than99.47%, and retrieval accuracy rate is close to100%. The related tasks of the project are successfully completed.