Abstract
We provide an introduction to Optical Music Recognition research along with a working tool for doing pre-processing of high quality scanned sheetm music to aid current commerical tools when the purpose is automated gathering of statistical information from large music corpora. This is implemented by using the Gamera Open Source framework for Document Image Analysis and uses a K-Nearest-Neighbor recognizer for recognizing and removing dynamic symbols like ``forte'' and ``pianisimo''. For text identification and removal different image projection and estimation techniques is used.
Tests was performed and results reported as precision and recall along with a weighted average know as the f-measure all concepts known from information retrival. In this report precision tells how many of the removed symbols actally should have been removed and showed a good result of 99%. Recall described how many of the total symbols we would like to remove that was actually found. For text the recall was around 84% and could possibly be improved. For dynamic removal the recall was found to be 98% whichs is quite reasonable. Qualitative tests to verify if preprocessing makes a difference was performed but it was not possible to show any beneficial effect.