GLSI Texture Descriptor Based on Complex Networks for Music Genre Classification

The texture classification of an image is related to an important musical attribute, the music genre. This relationship is depicted in the visual representation of the audio signal, called as spectrogram. In this paper, we propose a new Music Genre Classification (MGC) system that processes the spectrogram texture using the Gray Level and Structural Information (GLSI) descriptor, and represents the interconnection between the descriptor codes through complex networks. The GLSI descriptor is an improvement of the CLBP (Completed Local Binary Pattern) descriptor, which quantifies the texture of an image with three codes: signal (CLBP-S), magnitude (CLBP-M), and central (CLBP-C). By transforming the CLBP-C code, GLSI adds macro-structural information. The network nodes represent the descriptor codes, and the respective edges, the relationship according to the horizontal and vertical consecutive condition. We defined two representations for the nodes: 1) individual code node, obtaining the $G_{s},\ G_{m}$ and $G_{g}$ networks, and 2) triple code node, obtaining the $G_{smg}$ network. For the experimental stage, we used the GTZAN dataset, three types of spectrograms: conventional, mel-spectrogram and gammatonegram; and mining with network topological measures. For each type of spectrogram, we performed three experiments according to feature vector combinations, such as measures of: 1) $G_{s},\ G_{m}$ and $G_{g}$, 2) $G_{smg}$, and 3) all networks. In the machine learning stage, we used the ensemble classifier Bagging with Random Forest, and 10-fold cross-validation repeated 100 times. The experiment using all measures and all spectrograms revealed a satisfactory result, indicating that the MGC proposed is promising. We also propose a new equation to calculate the GLSI code, which proved to be much faster and with more intuitive encoding.

I. Introduction

Texture classification is an important task used by several digital image processing techniques, mainly because of its great utility in pattern recognition applications [1]. Specifically, in the field of computer artificial vision some applications can be found, such as classification of seismic textures [2] or recognition of traffic light [3], among others.