基于SVM的增量式音乐自动分类研究

发布时间：2018-06-19 22:15

本文选题：音乐分类 + 支持向量机　；参考：《南京邮电大学》2017年硕士论文

【摘要】：近年来,在互联网和多媒体信息技术快速发展的大环境下,人类已经步入前所未有的大数据时代,随之而来的就是海量的多媒体资源,其中就包含受众广泛的数字音乐。世界文化的多元化和大融合,使得音乐呈现出数量庞大、种类繁多、风格不一的特点。同时,对于听众来说,由于审美层次和文化背景的不同,并非所有风格的音乐都是满足他们喜好的音乐,这就要求音乐检索系统具备快速高效的分类能力。然而,传统的音乐检索系统的性能在很大程度上依赖于人工标注的训练样本。在当今这个数据量呈在线式增长的大数据时代,标注海量样本的做法显然是不现实的,依靠少量人工标注样本训练得到的分类器的泛化性能也会很差。因此,研究如何设计学习速度快、分类准确率高的音乐检索系统,具有非常重要的实用价值和现实意义。本文首先介绍了课题的研究背景,对音乐分类的国内外研究现状进行了分析;接着从历史训练样本集和新增训练样本集两个方面,对传统的基于SVM的增量学习算法进行了研究。在分析该算法在这两方面存在的不足的基础上,提出一种改进的基于SVM的增量式分类算法:针对历史训练样本集,引入包含更多分类信息的壳向量,代替传统算法所使用的支持向量;针对带有类别标签的新增训练样本集,提出基于KKT条件的差错控制策略,以此来选择最具代表性的样本参与训练。同时,为了使得算法更好地适应现实应用场景,针对大量没有类别标签的新增样本集,为了减少人工标注的成本,本文引入主动学习算法,提出一种综合考虑样本不确定性和多样性的价值样本选择策略,并将其应用于前文提出的增量式学习算法。仿真实验结果表明,算法在大大减少增量学习成本的同时,还保证了较高的分类准确率和较好的泛化性能。本文通过对传统的基于SVM的增量学习算法的分析,提出了相应的改进方案,并将其应用于音乐的自动分类,这对于未来关于音乐分类的研究工作具有一定的参考价值和现实意义。
[Abstract]:In recent years, with the rapid development of Internet and multimedia information technology, mankind has entered an unprecedented era of big data, followed by a large number of multimedia resources, including a wide range of digital music. The diversity and amalgamation of world culture make music present the characteristics of large quantity, various kinds and different styles. At the same time, for the audience, due to the different aesthetic level and cultural background, not all styles of music is to meet their preferences, which requires a rapid and efficient music retrieval system. However, the performance of traditional music retrieval systems depends largely on manual training samples. In the age of big data, where the amount of data is growing in line, the method of tagging massive samples is obviously unrealistic, and the generalization performance of classifier trained by a small number of manual labeled samples will be very poor. Therefore, it is of great practical value and practical significance to study how to design a music retrieval system with high learning speed and high classification accuracy. This paper first introduces the research background of the subject and analyzes the current research situation of music classification at home and abroad. Then the traditional incremental learning algorithm based on SVM is studied from the historical training sample set and the new training sample set. Based on the analysis of the shortcomings of this algorithm, an improved incremental classification algorithm based on SVM is proposed. For the historical training sample set, a shell vector containing more classification information is introduced. Instead of the support vector used in the traditional algorithm, an error control strategy based on KKT condition is proposed for the new training sample set with class label to select the most representative samples to participate in the training. At the same time, in order to better adapt the algorithm to the practical application scenarios, aiming at a large number of new sample sets without class labels, in order to reduce the cost of manual tagging, this paper introduces an active learning algorithm. A value sample selection strategy considering the uncertainty and diversity of samples is proposed and applied to the incremental learning algorithm proposed in this paper. The simulation results show that the algorithm not only reduces the incremental learning cost but also ensures higher classification accuracy and better generalization performance. Based on the analysis of the traditional incremental learning algorithm based on SVM, this paper puts forward the corresponding improvement scheme, and applies it to the automatic classification of music. This has certain reference value and practical significance for the future research on music classification.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3;TP18

【参考文献】