Research Of Optical Music Recognition

Optical Music Recognition (OMR) is the automatic recognition of printed music notation by image processing and pattern recognition. It scans original paper musical score into computer by scanner, and generates a bitmap of a scanned sheet of music. We can play it by the computer, or bring up in a music editor finally. OMR has many applications, for example, a soloist can practise with a computer playing an accompaniment; a musician can have a piece of music transposed to another key automatically; or an older edition of the music could be re-edited.This paper takes printed-image as the research target, and performs a deep research on OMR from rotation correction of staff line, location and adaptive removal of staff line, segmentation of music score symbol, feature extraction and recognition of music score symbol. New theories are proposed as follows.This paper proposes an adaptive staff line removal method for music score images. The proposed method consists of four main procedures, i.e., ROI (Region of Interest) selection, candidate point extraction, rotation correction and staff line removal. It selects a ROI firstly, then extracts candidate points in the selected ROI and estimates rotation angle with these candidate points finally. After rotation correction, the staff line width is estimated using horizontal projection and the staff lines are removed correspondingly. Extensive experiments show that the proposed method is robust for both printed and scanned music score images.We use horizontal and vertical projection in the segmentation of symbol for music score image. At first, for the whole image, we make use of some rules and then obtain a part of music score image with vertical projection. At second, we get region-music score image with horizontal projection again. At last, we obtain a symbol with vertical projection.In feature extraction of music score image, we add Fourier coefficient as the feature value. In recognition of music symbol, we adopt Back-Propagation Neural Network (BPNN) with non-linear clssificatioin, and good results have been achieved.