Audio Scene Recognition Based On Sample Re-balancing And Transfer Component Analysis

Audio scene recognition is to identify the corresponding scene informationthrough the analysis of the characteristics of the audio signal. At present, themain problem of audio scene recognition systems is the difference between the testsample distribution and the training sample distribution, in this case, the useof traditional learning methods can not be satisfied with the recognition rate. Inorder to solve this problem, this paper introduces the theory and method of transferlearning.This paper adopts Gaussian histogram as the scene feature. By the methods oftransfer learning to reduce the difference of sample distribution between training setand test set,then uses SVM to identity audio scene. This paper studies thecurrent main two kinds of transfer learning methods, transferring of instance andtransferring of feature. On transferring of instance，this paper researches the sampleselection bias,improves a type-independent correction of sample selection bias,which is the re-balancing by sample selection bias(RBSSB), and combined with thekernel density estimation theory, to remove the sample selection bias by re-selectingthe training samples, thus align the marginal probability distribution between thetraining and test samples. On transferring of feature, this paper researches thetransfer component analysis (TCA) and improves it, adds the optimization ofminimizing the within class scatter of the training set and maximizing the betweenclass scatter of the training set based on TCA, proposes the fisher discriminanttransfer component analysis (FDTCA), to make the transferred feature of differentkinds of scene has the better distinction. this paper also proposes the linear TCA, itgreatly reduces the calculation of transferring of feature.The experimental results show that, compared with the method without transferlearning, both RBSSB and TCA can improve the accuracy of the audio scenerecognition. Finally this paper merges them according to the characteristics of thetwo types of transfer learning, proposes the audio scene recognition based onRBSSB and TCA. The experiment proves that the recognition rate of the systemafter fusion is higher than using either methods.