基于生成对抗网络的音乐标签自动标注

针对如何快速有效地对音乐信息进行查询、检索和组织的问题，提出了一种基于生成对抗网络模型的多标签音乐自动标注系统.通过音乐自动语义标注技术，可以提高音乐检索系统的性能.利用LDA方法对音乐标签进行聚类以获取主题类别，再通过生成对抗网络，找到音乐的音频特征与语义特征之间的映射关系.应用于CAL500数据集的5次交叉验证实验结果表明，该方法的综合性能指标与现有方法相比有较大的提升.

Music auto-tagging based on generative adversarial networks

For the problem of how to query,retrieve,and organize music information quickly and efficiently,the performance of the music retrieval system can be improved through automatic music annotation technology.In this study,a multi-label music automatic annotation system based on generative adversarial networks(GANs) is proposed.The LDA model is used to cluster the music tags to obtain thematic categories,and then the mapping relationship between the audio features and the semantic features of the music is found by the generative adversarial network.For experimental verification,when the method proposed in this paper was applied to the CAL500 dataset in five cross-validation experiments,the comprehensive performance index of the method was greatly improved compared with existing methods.