CLBP Texture Descriptor in Multipartite Complex Network Configuration for Music Genre Classification

Abstract

Music genres define the characteristics that musical pieces must have to belong to a given class. These characteristics are reflected in the audio signal and, consequently, in the image that represents its spectral content: the spectrogram. In this paper, we propose a Music Genre Classification (MGC) system based on representation with complex networks of CLBP (Completed Local Binary Pattern) texture descriptor codes extracted from spectrograms: mel-spectrogram and gammatonegram. Complex networks were generated using CLBP codes in multipartite configuration: mono, bi, and tripartite networks; where the three node types are signal (CLBP-S), magnitude (CLBP-M) and central (CLBP-C) codes. The networks were mined using conventional, textural, and multipartite topological measures. In order to test the proposed MGC, we used the GTZAN dataset and defined several experiments using combinations of multipartite measures: 1) monopartite, 2) mono and bipartite, and 3) mono, bi and tripartite. All experiments were performed for each spectrogram individually and jointly. In the machine learning stage, we used the ensemble classifier Bagging with Random Forest, and 10-fold cross-validation repeated 100 times. As a main result, it was found that the bipartite measures related to CLBP-C decrease the performance, but the tripartites increased it. Moreover, in most experiments using only gammatonogram the performance was better. Consequently, the experiment using tripartite measures extracted from the gammatonegram revealed a satisfactory result, indicating that the proposed MGC is promising.