The Research Of Music Genre Classification System Based On Auditory Images

With the popularity of internet, how to select the information satisfies needs ofusers from tens of thousands of music on the internet, became the issue requiringsolution. In this way, the content-based music information retrieval technologybecomes crucial in the field of information retrieval technology. Music genreclassification becomes one of the hot spots for domestic and international experts andscholars to research in recent years.In this paper, we introduce the auditory image model into the field of musicgenre classification. The auditory image model converted the one-dimensional audiosignal into two-dimensional auditory images by simulating the human ear cochlearstructures using a series of mathematical expressions. And then, extract texturefeatures from different genre images by the methods of Scale Invariant FeatureTransform (SIFT) and spatial pyramid matching (SPM). SIFT can extract the spatialposition, size and gradient orientation information from each piex. SPM can integratethe local feature into the global features in different levels. The feature vector basedon auditory image is high dimensional and sparse. The statistical pattern recognitionwas chosen to complete the automatic musical genres classification, such as KNNwhere k with different values and SVM with different kenel functions.The linearkernel support vector machine got the best accuracy for classification. Experimentalresults showed that the genre classification accuracy based on the auditory images canbe15%higher than the Mel-frequency cepstral coefficients (MFCC) which was alsobased on the cochlear structure of the human ear, and almost6%higher than timeralfeature added "Diffusion Maps". The classification in this paper is also better than theclassification based on spectrogram of MFCC and SIFT.