Enhanced Feature Summarizing for Effective Cover Song Identification

Self-similarity analysis-based feature summarizing technique (SuCo) was proposed recently to improve the time and memory efficiency of Cover Song Identification (CSI). In this paper, both the feature summarizing and the cross-similarity calculating strategies of the SuCo model are modified as follows to enhance its identification accuracy. At the feature summarizing stage, first, the Hubness Reduction (HR) strategy is adopted to reduce the possible `Hubness' phenomenon existing in the feature subsequence community, which may affect the retrieval effectiveness. Then, the Network Enhancement (NE) technique, which was originally proposed in biology to improve gene-function prediction accuracy, is introduced to reduce the noise in the self-similarity network caused by the limitation of feature extraction and similarity measuring, and the inherent musical and acoustic variations. At the cross-similarity calculating stage, first, the summarized representative feature subsequences of the reference are concatenated to obtain its combined representative feature. Then, considering that the nonlinear recurrence property is important for describing the melody perception-based similarity, Qmax is adopted to measure the similarity between the combined representative feature of the reference and the unsummarized feature sequence of the query. Extensive experiments carried out on four open CSI datasets with 5 types of features and 2 kinds of representative feature subsequence choosing methods verify that: i) The proposed scheme outperforms the SuCo model in retrieval effectiveness. ii) Each of the above modifications contributes to the performance enhancement of the proposed scheme. iii) The proposed scheme achieves high generalization.