Coordinating knowledge to improve optical music recognition

Optical Music Recognition (OMR) is the process of automatically processing and understanding an image of a music score. This process involves various distinct phases to transform the image into primitive shapes, musical objects, and ultimately into a syntactic model representing the music's semantics. In general, OMR systems have performed these tasks in a linear sequence, so that the output of one component is the input to the next. However, this means that processing errors that occur in one of the tasks propagate through the system, and often when the error is eventually detected it is too late to reconsider the decisions leading to the incorrect classification or information. This thesis describes how OMR can be improved by modifying the recognition process from a sequence of linear tasks to a collection of modules that coordinate the information extracted from the data. Methods for_ data representation and controlling the system's flow of execution are investigated, and a practical implementation of such a system is described. This system has a message-passing design for providing contextual information from one module to another, such as suggesting possible classifications for an object. These messages are used to aid decision-making and to correct faulty decisions. This helps the system to adapt to a particular score while processing the image, increasing accuracy. This system is designed to aid in the research and evaluation of algorithms to achieve the above aims; therefore it is straightforward to modify various aspects of the system's behaviour, such as adding support for different music symbols. Examining the implemented system's behaviour clearly shows that this coordinated approach can correct many errors and can even identify some objects by only using syntactic information, based on the surrounding objects