Research On Printed Score Recognition Method Based On Deep Learning

Optical Music Recognition(OMR)is an important way to realize the digitization of musical score images.It has wide application prospects in computer music,digital music library,computer-assisted music teaching and other fields.Recognition of musical symbols is an important research content of optical score recognition.At present,there are some difficulties in music symbol recognition: the structure of the music score is very complicated,the size of the music symbol is extremely small,and some symbols are too dense(chords),etc.,which leads to low recognition accuracy of the music symbol.With the rapid development of artificial intelligence technology,deep learning algorithms have been tried and achieved preliminary results in the field of OMR,which has opened up a new path for OMR research.This article takes printed musical scores as the research object and researches the recognition of musical symbols based on deep learning algorithms.The specific research content is as follows:Firstly,for the problem that the current OMR data set has incomplete labeling information and can not recognize the duration and pitch of the note,study and generate a dataset containing pitch and duration tags.First of all,by analyzing the MXL(Music XML)file,the score image and corresponding label data required for model training are obtained;then,the duration label of the note is obtained by connecting the information such as the beam,the dot,and the flag,and the pitch is parsed from the position information of the note Labels;Finally,data enhancement methods such as noise and random cropping are added to increase data diversity.Secondly,to solve the problems of OMR recognition process is complicated and recognition accuracy is not high,the end-to-end printed music score recognition model based on YOLOv3 is studied.The model takes the entire score image as input,and directly outputs the symbol duration,pitch and classes.The model adopts multi-task learning,learning the classification tasks of pitch and duration and the regression tasks of note coordinates,which improves the recognition speed of the model.The model of this paper is verified by the score image testset.The experimental results show that the accuracy of note recognition is high,which can reach the accuracy of duration value of 0.96 and the accuracy of pitch of 0.98,which is higher than theaccuracy of note recognition of other score recognition models.Thirdly,in view of the difficulty of detecting polyphony(chords)in the score image,polyphonic score image recognition based on Fine-Tuning technology is studied.First,manually label an appropriate amount of score images as a polyphonic recognition data set;then add additional regression branches for specific symbols(tuplets)to improve the recognition accuracy of tuplets.Finally,combined with the neural network fine-tuning technology,the end-to-end printed music score recognition model based on YOLOv3 is fine-tuned to improve the accuracy of polyphonic music score recognition.The experimental results show that the recognition accuracy of polyphony and tuplet is greatly improved while keeping the symbol recognition accuracy unchanged.