Music genre classification using visual features with feature selection

This paper describes the results obtained in a set of experiments performed making feature selection using Genetic Algorithms. The objectives were to improve the recognition rates or to reduce the feature vector dimensionality in music genre recognition task. Firstly, music pieces taken from Latin Music Database were decomposed into 3 different sub-signals, each one with 10 seconds from beginning, middle and end parts of the original music. After binding these three parts, the 30 seconds music signals were converted into spectrograms. Considering a local feature extraction rather than a global one, a zoning mechanism is performed. In the experiments, we used the Local Binary Pattern texture descriptor in the feature extraction step, Genetic Algorithms for feature selection and Support Vector Machine to perform the classification. In short, obtained results have shown that it's possible to get an important dimensionality reduction in feature sets without great changes in recognition rates.