Abstract
Nowadays records, radio, television and the internet spread music more widely than ever before, and an overwhelming number of musical works are available to us. During the last decades, a great interest in converting music scores into a computer-readable format has arisen, and with this the field of Optical Music Recognition.
Optical Music Recognition (OMR) is the name of systems for music score recognition, and is similar to Optical Character Recognition (OCR) except that it is used to recognize musical symbols instead of letters.
OMR systems try to automatically recognize the main musical objects of a scanned music score and convert them into a suitable electronic format, such as a MIDI file, an audio waveform or ABC Notation.
The advantage of such a digital format, compared to retaining the whole image of a music score, is that only the semantics of music are stored, that is notes, pitches and durations, contextual information and other relevant information. This way much computer space is saved, and at the same time scores can be printed over and over again, without loss of quality, and they can be edited and played on a computer \citep{Vieira01}. OMR may also be used for educational reasons - to convert scores into Braille code for blind people, to generate customized version of music exercises etc. In addition, this technology can be used to index and collect scores in databases.
Today, there are a number of on-line databases containing digital sheet music, making music easily available for everyone, free of charge.
The earliest attempts at OMR were made in the early 1970's. During the last decades, OMR has been especially active, and there are currently a number of commercially available packages. The first commercial products came in the early 90's. However, in most cases these systems operate properly only with well-scanned documents of high quality. When it comes to precision and reliability, none of the commercial OMR systems solve the problem in a satisfactory way.
The aim of this thesis is to study various existing OMR approaches and suggest novel methods, or modifications/improvements of current algorithms. The first stages of the process is prioritized, and we limit to concentrate on identifying the main musical symbols, essential for playing the melody, while text, slurs, staff numbering etc. are ignored by our program. The last part of an OMR program usually consists of correcting classification errors by introducing musical rules. In this thesis, this is only applied to correct wrongly classified pitched for accidentals.