Research On Incremental Music Automatic Classification Based On SVM

In recent years,under the environment of Internet and multimedia information technology developing rapidly,mankind has entered the unprecedented era of big data,along with massive multimedia resources,including the digital music with a broad audience.The diversity and integration of the world culture make the music show the characteristics of large quantity,wide variety and different styles.At the same time,for the audience,due to the different levels of aesthetic and cultural background,not all the style of music is to meet their demands,which requires the music retrieval system with a fast and efficient classification ability.However,the performance of the traditional music retrieval system is largely dependent on the artificially labeled training samples.In the era of big data with the data showing a online growth,the approach of labeling massive samples for training is clearly not realistic,and relying on a small number of manually labeled samples,the obtained classifier by training would has a poor generalization performance.Therefore,it is of great practical value and realistic significance to study how to design a music retrieval system with fast learning speed and high classification accuracy.Firstly,this thesis introduces the research background,and analyzes the present situation of the research on music classification at home and abroad.Then,from the two aspects of former training sample set and newly added training set,this thesis do some research about the traditional incremental learning algorithm based on SVM.By analyzing the shortcomings of algorithm on the two aspects,this thesis proposes an improved incremental classification algorithm based on SVM:to the former training sample set,instead of using the traditional support vector,introducing the hull vectors,which contains more classification information;to the incremental sample set with category label,proposing the error control strategy based on KKT conditions to choose the most representative samples used to training.At the same time,in order to make the algorithm better adapt to the realistic application scenarios and reduce the cost of manual annotating to the large number of newly added samples without labels,introducing the active learning algorithm,presenting a kind of selection strategy for valuable samples with a comprehensive account for the uncertainty and diversity of new added sample,and applying it to the proposed incremental learning algorithm before.The simulation experiment results show that the algorithm can not only reduce the cost of incremental learning,but also guarantee the high classification accuracy and good generalization performance.By the analysis about the traditional incremental learning algorithm based on SVM,this thesis puts forward the corresponding improvement schemes,and applys them on the automatic music classification,which has a certain reference value and practical significance for the future research work on music classification.