基于DAG-LDM的手写音乐符号分类

Handwritten music symbol classification based on DAG-LDM

光学乐谱识别(optical music recognition,OMR)是将纸质的音乐乐谱转化为计算机可以读取的格式。其中音乐符号分类是乐谱识别的重要一个步骤。本文并提出了一种基于DAG-LDM的手写音乐符号分类方法,该方法利用DAG有向无环图结构将二类分类器LDM拓展到多类,对于N类音乐符号,需要训练N(N-1)/2个分类器,并依照音乐符号的特征合理排序,防止DAG结构的错误累积效应,用这种DAG结构解决多类分类问题效率高,速度快。在分类器方面,相比支持向量机(support vector machine,SVM)只优化了最小间隔,DAG-LDM还优化了间隔的分布,这更符合音乐符号的样本分布特性,并且具有较强的抗噪性能。本文将这种新DAG_LDM音乐符号分类算法与几种主流的方法进行对比测试,测试结果显示本文提出的新算法对手写音乐符号分类具有更高的分类准确率。且本文提出的算法不仅仅适用于音乐符号识别,还可以用于其他的多类分类问题中。

The task for an optical music recognition（ OMR） system is to digitalize and transform the original music sheets into a machine-readable format. The music symbol classification is the most important part of OMR. In this paper,a new classifier based on directed acyclic graph-large margin distribution machine（ DAG-LDM） is presented. The DAG-LDM is extended from the binary classifier LDM（ Large margin Distribution Machine） using the DAG. For a problem of N classes music symbols,N（ N- 1） /2 classifiers are trained. The order of the classes of music symbols is organized to avoid the error accumulation effect. The DAG model is more efficient on dealing with multiclass problems. The SVM considers the minimum margin to get the optimal decision. In DAG-LDM,not only the margin mean is maximized,but also the margin variance is minimized. The algorithms are compared with the same data sets,which are handwritten images. The tests show that the proposed method provides superior classification capability and achieves much higher classification accuracy than that of the state of art algorithms such as the support vector machine（ SVM） and neural network（ NN）. The proposed method is not only suitable for music symbol classification,but applicable to other multi-class problems as well.