Desenvolvimento de um reconhecedor óptico de símbolos musicais em partituras monofônicas ocidentais

Abstract

Optical music recognition is the area of computing that studies ways to identify and interpret musical symbol images of a musical score. From the interpretation of these symbols, it is possible to extend the study into the field of analysis or application. There are projects that use optical music recognition to convert visual information from scores into MIDI format music files. On the other hand, there are applications that convert this information into a more structured format, such as MusicXML, allowing a broader analysis work, even for identification of composition patterns. This academic project aimed to develop an optical recognizer for western monophonic musical scores, focusing on a few known musical symbols, using image processing and classification techniques, encoding the final representation into MIDI format. For image preprocessing, the global and Otsu binarization methods, the Gaussian filter and morphological erosion and dilation operations were used. For the classification of the symbols, it was adopted a model trained through the convolutional neural network techniques in conjunction with a recurrent neural network, using the PRIMUS dataset. For prediction, in addition to the PRIMUS dataset, proprietary staves were used for further validation and classification tests. The own staves were generated using the Noteflight web system. The prototype returned the best results using the Otsu binarization in conjunction with the Gaussian filter with a 3x3 mask, and the most suitable rotation parameter for the alignment of the staves was 0.25 degrees. The total percentage of correctly classified symbols reached 97.43% for the selected images from the PRIMUS dataset and 77.38% for the generated own images. [resumo fornecido pelo autor]