Music Genre Classification Based On Acoustic Feature Weighted Fuzzy Support Vector Machine

Music genre as an important basis of music retrieval,the study on its classification is of great significance.Preprocessing stage work mainly includes the audio signal processing and acoustic feature selection,there has been a set of mature standards for feature extraction that can extract hundreds of dimensions of music features.However,there is still no consensus on feature selection methods.In the classification stage,SVM is often used in music genre classification due to its good generalization ability,but its sensitivity to noise At the same time,because of the fuzziness of music genre itself,the previous output mode of single category will cause the loss of information.Music genre is taken as the object in this paper,and two research points of improving the classification accuracy of music genre single category and improving the genre information integrity of music classification are respectively carried out.The research content is as follows:(1)A fuzzy support vector machine method based on Relief-F feature weighting is proposed for music genre classification.The weight of each music feature is calculated by using the Relief-F algorithm,which has a good explanatory power for the selected feature.Fuzzy support vector machine is selected as the classifier,and a new membership function is constructed based on the weight calculated by relief-f to reduce the impact of noise on classification,taking into account the different impact sizes of different weight features on classification.Directed acyclic graph(DAG)fuzzy support vector machine(FSVM)is adopted for the combination of multiple classifications,because it has no problem of inseparable regions and has high classification efficiency.(2)A multi-category probability classification method is proposed to improve the genre information integrity of music classification.The fuzzy support vector machine(FSVM)is also used in the classifier.By comparing three kinds of multi-classification methods,the non-separable region formed by the combination of 1-v-r classifier is adopted to identify music samples with diversified schools.Finally,weighted Euclidean distance is used to calculate the probability of each target category.According to the different needs of different business scenarios,this paper proposes the method to improve the accuracy of single category classification of music genre with good effect and the way to improve the genre information integrity of music classification,and applies it to the data set of ISMIS2011 music genre classification competition.The effectiveness of the first research site was verified by comparing it with the best results of the competition.At the same time,it is found that the target category of genre diversity music sample discovered by the traditional classification method is the same as the target category of genre diversity music sample discovered by the multi-category probability classification method in this paper,which verifies the effectiveness of the second research point.