Improved similarity fusion scheme for cover song identification
Improved similarity fusion scheme for cover song identification
复制标题
DOI：
10.1049/el.2018.6461
复制
发表时间：
2018-10
影响因子：
1.1
通讯作者：
Yanlan Fan;Ning Chen
中科院分区：
文献类型：
--
作者：
Yanlan Fan;Ning Chen
提出了一种基于非线性图融合和张量积图扩散的翻唱歌曲识别(CSI)方案，改进了作者以前提出的基于相似度融合的CSI方案。首先，从音轨中分别提取和声级数、旋律演变和基于节奏的描述符。其次，采用相似网络融合的方法对基于两类描述符得到的相似度图进行融合，充分利用了它们之间的共性和互补性。最后，利用融合后的相似图所包含的流形结构，对融合后的相似图进行TPGS扩散，以进一步提高算法的性能。实验结果表明，该方法在识别精度和聚类性能方面均优于以前提出的方法。
A Cover Song Identification (CSI) scheme based on non-linear graph fusion and Tensor Product Graphs (TPGs) diffusion is proposed as an improvement to the authors' previously proposed similarity fusion-based CSI scheme. First, the harmonic progression, melody evolution, and rhythm-based descriptors are extracted from the track, respectively. Next, Similarity Network Fusion is adopted to fuse the similarity graphs obtained based on two types of descriptors to take full use of the common as well as complementary properties between them. Finally, TPGs diffusion is performed on the obtained fused similarity graphs to take advantage of the manifold structure contained in them to improve the performance, further. Experimental results demonstrate the superiority of the proposed scheme over their previously proposed one, in terms of identification accuracy and clustering performance.