This end-of-degree work aims to study the general situation of optical music recognition with the aim of trying to improve an already existing tool for real-time symbol labelling for musicologists. From four fragments contained in a 17th century manuscript corpus and using image processing and automatic learning techniques, an incremental classification system has been designed that is capable of improving its success rate. For this purpose, the use of different algorithms has been studied, comparing their strengths and weaknesses. In addition, the creation and use of different initial training sets has been assessed, allowing the tool to start from a minimum error rate. All these experiments have managed to find a way to decrease both the error rate and the classification time, forming an efficient system of automatic labelling of musical symbols.
Technical report (in spanish):