Font Size: a A A

Research On Perceptual Audio Hashing Functions For Music Emotion Retrieve

Posted on:2015-11-01Degree:MasterType:Thesis
Country:ChinaCandidate:Z G DuanFull Text:PDF
GTID:2428330488499884Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
In music information retrieval and recommendations,how music emotion efficient accurate identification and retrieval,subject of great research value that intelligent music system further developing.Emotions and emotional reactions that people listen to music can be quantified into a unified model and people will retrieval music based on music emotion.The first problem of research on music emotion retrieval is get a lot of data with emotional information marked.This paper considers the use of Web2.0 user's emotional evaluation information for music,build music emotion database,through to the realization of music emotion retrival.Using information extraction methods,have been able to get a text message from the emotional imformation of music,but because digital audio is not unique,so how to mapping the audio to the laber based on perception content,is the key scientific problems to be solved to build musical emotion corpus.Based on this,fast perceptual audio hash function try to solve it,focusing on the need to improve its time performance.In this paper,two efficient audio perceptual hashing methods is being try:(1)Using vector projection and random search strategy,starting with a large number of training data obtained in the search of suitable projection vector,then the high-dimensional perceptual features projected onto a one-dimensional space,and coding.Experimental results show that the method not only without reducing the distinction between the hash vectors and robustness of the premise,but also improve the time performance compared to traditional methods has greatly.(2)Based on deep belief networks,using DNN-Bottleneck framework to extracte robust perceptual features from the audio data,and then encoding.Experimental results demonstrate the feasibility of this method,and compared to VP-FFT method,it's distinguish and robustness has improved.Further analysis revealed that,VP-FFT method and monolayer NN-Bottleneck-FFT method hash the same mathematical describe,it is provides a theoretical explanation reference for VP-FFT method.
Keywords/Search Tags:Music Emotion Retrieval, Perceptual Audio Hashing Functions, Vector Projection, Strategy of Random Search, Deep Belief Networks, Bottleneck Feature
PDF Full Text Request
Related items