Audio Index Based On LSH Distance And Retrieval System

It is an important research of how to design high-dimensional indexing structure incontent-based multimedia information retrieval. The indexing based on spatial access methodwill lead to“dimension disaster”;The indexing based on metric method has been used widelybecause it works more effective. HCT（Hierarchical Cellular Tree）is a index tree based onmetric method which grows dynamically and it proposes a data organization methodincluding a fast search method. When using HCT, the structure of index items and thedistance expression between items must be design newly. In addition, currently plenty ofretrieval systems ignore the attribute of time-sequence in audio retrieve which lead to a lessexact retrieval result.Addressing the above problems, this thesis designs the audio index construction andretrieval method. The calculating of distance between feature vector cost most of the timewhen constructing a CHT index tree. In order to reduce the time cost, in this thesis, the use ofdimensionality reduction is LSH（Locality Sensitive Hashing）. The main work about thisthesis is listed below:1. Creating the audio index. Firstly, filtering the mute segment and segment audio intoclips, then Extracting feature parameters constitute a feature vector and normalizedfor each component. Structuring audio clips of CHT index tree and using Euclideandistance to describe the distance between two index clip.Finally, generating audioindexing documents in the HCT software platform.2. Through the LSH technology to optimize the index system, improve the retrievalspeed. Because of problem of big calculation of the computation of high dimensionalfeature vector, the technology of LSH mapping high dimensional feature vector tointeger low dimensional space, representing Index fragments with less data. Distancecalculation between clips is greatly reduced, so that retrieval time is also greatlyreduced.3. The input query audio is segmented query, this paper proposes a ordering strategybased on Comprehensive score.Segmented audio through HCT fast retrieval ofcandidate index segment sets, according to the time sequence of the candidate set.Screening out the continuous in time segments of candidate target result, to meet therequirements of the target results were scored to determine the degree of correlation,The ultimate goal of results by interested degree ordering. 4. Perfect man-machine interaction design, To provide a friendly user experience.Retrieval system based on B/S architecture, Support local upload sample and selectonline sample retrieval, Support the retrieval object automatic orientation broadcast,Support video content browsing based on audio sound type.