Aplicação de redes neurais artificiais para o reconhecimento de elementos de partituras musicais

Abstract

The importance of music for humanity is notorious, as it has influenced different aspects of societies for centuries. In addition to its cultural value, it can contribute to cognitive development and is related to a market of great financial importance. However, the diversity, volume and complexity of musical records make maintaining and manipulating sheet music a challenge. To digitize them, manual processes can be time-consuming and prone to errors. Therefore, the use of automatic sheet music recognition methods is a pertinent objective. In this context, the present work aims to explore the application of neural networks in classifying elements of musical scores. To perform this task, multilayer perceptron networks and convolutional neural networks are used. Several configurations are tested to evaluate the influence of the parameters. Several configurations were tested to assess the influence of parameters such as the number of samples per batch, number of training epochs, kernel dimensions, and number of layers, considering convolutional, dense, dropout, and normalization layers. In this study, it was achieved with the best network for figure classification, an accuracy of 90.8% for the test set. Whereas for note classification, the top-performing network reached 85.2%. In both cases, the best-performing networks were convolutional. On the other hand, the MLP network did not exhibit effective performance. The correct classification of sheet music elements lays the foundation for the complete digitization of these musical representations. Furthermore, the neural networks discussed can be used in other applications, such as pattern recognition in images and sounds, or optimization problems.