Application of Convolutional Neural Network-based Note Recognition and Analysis Technique for Piano Performance in Basic Piano Teaching

Abstract

Optical sheet music recognition converts traditional paper sheet music into electronic sheet music, which is convenient for players to learn and practice playing. In this paper, taking the piano notes as the research object, combining the training needs of piano playing technique with the features of piano music and its recognition characteristics, the residual network is used to improve the learning ability of the network and the convolutional layer is adjusted to be a residual depth-separable convolutional network. The recurrent layer is fed with SRU to accelerate the learning speed of the network and accelerate the note classification. The transcription layer uses Focal Loss to deal with note overfitting samples, forming a lightweight note recognition method based on improved CRNN.The NSynth Dataset dataset tests show that the accuracy of the three sub-networks reaches a high level when the model is trained up to 17 times, and there is no overfitting problem in the trained network model.The addition of the SRU module reduces the model training time and The optimal sequence error rate of the whole model for the deformed semantic sheet music is 32.89%. The improved CRNN network can improve students' performance in piano playing test by applying it to piano intelligent assisted playing training.