Design And Implementation Of Music Sheet Recognition Software

Musical scores,originally printed in paper,are used to record music with symbols.With the development of electronic technologies,they can be scanned or photographed into images,thus the problems in storage,copy and distribution of paper scores are solved.The booming of computer and Internet technologies has brought various music applications,and music files that can be understood by computers provide fundamental data for these applications.In this thesis,music score images,as the research object,are transformed into MIDI files with the algorithms of image processing and recognition,and then a musical score recognition software is implemented.The contribution of this thesis is as follows:Firstly,based on the analysis of the requirements of the music recognition software,the system framework and its key technologies are proposed,including image preprocessing and correction,staff line positioning and removal,note segmentation and recognition,etc.,the software development environment is established and data set for convolution neural network training is prepared.Secondly,the above mentioned key technologies are researched:(1)Bilateral filter and image enhancement algorithms are used to denoise the image,and an improved Hough transform based on edge band is proposed to correct image rotation;(2)Stable path is calculated to locate staff lines in despite of their deformation;then an algorithm based on local run length histogram is proposed to detect and remove staff lines,which effectively protect the completeness of notes when staff lines are removed;(3)The algorithm based on connected domain and hierarchical decomposition is proposed for note segmentation.A satisfying result is achieved with coarse and refined segmentation combined with the rules in note size;(4)Convolution neural network in deep learning combined with traditional template matching algorithm is used for note recognition,which overcomes the limitations of traditional recognition technology in image adaptability and robustness.Finally,a musical score recognition software based on the above algorithms is implemented in vs2013 with opencv3.0 and tensorflowl.8 C++plug-in,and a human-computer interface is designed based on QT.The effectiveness of the above algorithms is verified by experiments and result analysis.