Neural Networks for Optical Music Recognition

This thesis consideres the problem of optical music recognition from images to text using Artificial inteligence and neural networks. I have choosed particularly the field of printed polyphonic music (more notes and voices at the same time). The goal of this thesis is to create a model capable of recognising complex notations and its accuracy compare with previous literature and other known models. I solved the chosen problem by utilizing the Vision Transformer architecture, where I tested several network variants to find the most powerful one. And by creating a new dataset with polyphonic music. The work presents the process of creating the dataset by synthesizing images from MusicXML format using the MuseScore program. The most successful variant of the Vision Transformer architecture achieves an error rate of only 7.86 %, which is very promising for further development and utilization. The main finding is that the architecture has the potential to dominate in this field, just as it does in other areas of research, and there is a functional solution for the specific task of polyphonic music notation recognition, which has been only up for a debate until now