Research Of The Algorithm Of Music Score Correction And Generation Of Universal Music Score Dataset

Optical Music Recognition(OMR)is a field that researches how to transform music photos into computer-readable format files.The content of music image can be a single music symbol or a music score paper.In this paper,the research object is the music score.OMR enables that people can complete the digitalization of paper music score by means of photographing and scanning without the professional music theory,so OMR greatly reduces the cost of digitalization of music score.With the development of deep learning,some researches of OMR based on deep learning have been put forward.But existing open datasets cannot meet the data needs of OMR research based on deep learning.Meanwhile,with the popularity of mobile devices,the cost of music score images collection is lower,but it also leads that the quality of images can’t be guaranteed.In OMR system,the low quantity of input images will increase the difficulty of subsequent recognition.In this paper,the two research tasks are to solve the above problems.In first task,aiming to the deep learning dataset problem,this paper proposes a construction scheme for synthesizing diverse music notes datasets.The whole process is divided into three steps.In the first step,the project modifies the MusicXML files of simple music scores in order to increase the note types.And then,these MusicXML files are converted into music score images by open API.Finally we design content labels for music score images.Meanwhile,this paper realizes synthetic data samples based on deep learning.We perform style transfer CycleGAN model for synthetic images.The transferring can make the synthetic images to obtain real photo texture while keeping the content unchanged.In first task,aiming to the deep learning dataset problem,the project builds a universal music score dataset with a wide range of notes in the method of artificial synthesis.The whole process is divided into three steps.In the first step,the project modifies the MusicXML files of simple music scores in order to increase the note types.And then,these MusicXML files are converted into music score images by open API.Finally the project takes the style transfer operations for these music score images in CycleGAN model in order to make these synthetic images similar to the real photos.The dataset can be used to train neural network model for various music scores recognition tasks because of its features of the wide range of notes,containing chord symbols and similar to the real photos.The second research task is carried out from the term of improving the quality of music score images.The project restores the irregular image to a regular shape by correcting the irregular deformation in music score images in order to reduce the difficulty of OMR and finally improve the recognition accuracy.Meanwhile,the rectification process has a gain effect on OMR systems based on both traditional algorithm and deep learning algorithm.The rectification process is based on deep learning,so the project starts with the synthesis of dataset.Firstly,the deformation grids are synthesized,the regular music score images are mapped into new deformation images as the model training dataset according to these deformation grids.Then a patch-based CNN model predicts the deformation grids for blocks of a whole score image.In the next step,these predicting results,deformation sub grids,are stitched as one grid.Finally,the irregular images are backward mapped into final rectification results.After backward mapping,the whole rectification process is completed.