The thesis presents two approaches to optical music recognition: the traditional one, in which the problem is solved by the strategy of "divide and rule", and the newer, holistic approach, in which the problem is solved with a deep neural network. Both are described and compared in more detail, and their implementation, optimization and achieved results are also presented. The collections of music incipits PrIMuS and CorPus, with which the models were evaluated, are described. Two different accommodation systems in this field (paid and open source) are also presented. Finally, the results of the models created as part of the diploma thesis and the results of existing systems are compared. The best results were achieved by the method with a deep neural network, while the optimized traditional method could be used with precisely defined or limited problem.