Research On Printed Score Recognition Technology Based On Deep Learning

Optical Music Recognition（OMR）is an important way to realize the digitization of music score images,and has a wide application prospect in the fields of computer music,digital music library,and computer-aided music teaching.The recognition of music symbols is an important research content of optical music recognition.The complex structure of music scores,the small size of music symbols and the overly dense arrangement of some symbols lead to the low recognition accuracy of music symbols.With the rapid development of artificial intelligence technology,deep learning has received extensive attention and achieved preliminary results in the field of OMR,which provides a new research path for the study of OMR.In this regard,this paper takes printed music scores as the research object and uses deep learning technology to study the music symbol recognition method for music scores.The specific research contents are as follows:Firstly,the dataset is pre-processed.Since the notes in the Deep Scores dataset lacked duration labels,in this paper,in order to directly predict the note duration,we adopted the manual labeling data to add duration to each note in the score,and the note duration can be used as the training label when training the music symbol detection model;secondly,we added noise,fuzzy processing,elastic deformation and color transformation to the original score to expand the The data set can be expanded by adding noise,fuzzing,elasticity and color transformations to the original score.Secondly,based on the end-to-end music symbol recognition research.Researched the Musical Symbol Recognition Model based on Center Net.The model directly inputs the entire score image and adopts a multi-task learning mechanism.At the same time,it trains the classification task of musical symbols,duration and the regression task of musical symbol coordinates.In order to improve the mean accuracy（AP）of symbols of different sizes,multi-scale prediction technology is incorporated on the basis of the original model,that is,independent training and prediction are performed on three feature maps of different scales,and then non-maximum suppression（NMS）After the algorithm is processed,the final prediction result is obtained.The experimental results show that the recognition accuracy of conventional music symbols is high,and the duration accuracy rate can reach 97%.Thirdly,the study of music symbol recognition based on staff line removal.Firstly,the staff line removal model based on the semantic segmentation model U-Net is studied,which adopts the multi-scale structure in Inceptionv3 to obtain the rich semantic information of the image,and replaces the pooling layer with dilated convolution to prevent the loss of image detail information,after which the Dropout layer is introduced to prevent the overfitting phenomenon and also to reduce the training time;In the music symbol recognition stage,the scores without staff line are used as the data set to train the music symbol recognition model and compared with the end-to-end method.The experimental results show that for polyphonic scores with complex layout,the removal of the staff line is beneficial to improve the accuracy of dense symbols（chords）as well as small target symbols（rests）.