Music Score Recognition System Based On Embedded Platform

In order to automatically and intelligently read the paper music and play the corresponding music to improve the efficiency of music teaching,and promote the application of deep learning technology in optical music recognition,this paper mainly develops a set of music score image recognition system based on embedded platform,Studies and realizes the music image recognition method based on depth learning technology,and realizes the automation and intelligence of music image recognition.The music recognition system based on embedded platform includes two parts:embedded platform and musical score image recognition algorithm.The embedded platform mainly provides software running environment and hardware support for the digitization of paper score,and the music score recognition algorithm is based on the depth learning technology to digitize the music score image.Embedded platform is based on NVIDIA Jetson TX2 development board as the core hardware,with Logitech C270 camera,deep learning server monitor and speaker,constitute the image recognition system hardware.On this basis,build a system software environment based on Linux system and Caffe deep learning framework.In the system input,using OpenCv open source visual library to call the camera,and the collected digital music image size conversion,brightness conversion and other processing.Then based on the core algorithm of deep learning to identify digital images and generate audio files.At the system output,the audio files are decoded by timidity decoding,and the recognition result is played back.Music image recognition algorithms include spectral positioning,note detection and pronunciation of the three elements of the analysis.First of allHuff transform is used to detect spectral lines,and then fused with FAST corner detection,image pyramid,concatenated convolutional neural networks and other image processing techniques to detect notes,and then the image segmentation technique is used to segment the notes of the pronunciation primitives.After segmentation The primitives are analyzed and combined to reconstruct the independent note information.The innovation of this thesis is as follows:(1)A note dataset with different lighting conditions and different sharpness conditions was set up according to printed notes.According to a large number of musical notes in domestic folk songs and foreign famous songs,scores lines,independent notes and pronunciation primitives Level database that supports training and testing of deep learning algorithm models.(2)For improving the deficiencies and limitations of the traditional optical music recognition technology in adaptability and robustness of image data,the traditional image processing method is combined with the convolution neural network in depth learning to improve the effect of the music image recognition and the accuracy reached 84.6%,the recall rate reached 94.8%.(3)Integrate and optimize each function module of music image recognition system through embedded platform,and design concise human-computer interaction interface to operate,realize the integration,automation and intelligence of paper music score to the conversion of electronic audio.