Research on Handwritten Note Recognition in Digital Music Classroom Based on Deep Learning

Abstract

Music is an indispensable subject in quality education, which plays an important role in improving students’ overall quality. Traditional music teaching is mainly a one-way teaching led by teachers. The teaching style is monotonous and the teaching resources are lacking. It is in sharp contrast with diversified music learning, which affects students’ mastery of basic music skills. The research of this paper is mainly based on the metaphor of “paper and pen”, and the user-centered natural interaction method is used as the design idea to identify the handwritten notes in music teaching in order to provide a natural and efficient teaching environment for music education. Handwritten note recognition draws on the idea of metric learning. Based on the deep Gaussian process model, a non-parametric model, a deep Gaussian matching network for small batch handwritten note recognition is proposed. The framework can adaptively learn a deep structure that can effectively map the labeled support set and unlabeled samples to its label, while avoiding overfitting due to insufficient training data. In the training stage of the deep Gaussian process model, the standardized flow method is used to construct a flexible variational distribution, which improves the quality of inference. Gaussian Processes (GP) are type of supervised learning system that can be used to solve problems like regression and probabilistic classification. Gaussian processes have the following advantages: The forecast generalizes the data from the observations. And when sparse the Gaussian model to reduce the amount of calculation, the optimal k-means method is used to find false points. Experiments were carried out on the handwritten note data set collected in the digital music classroom. The experimental results show that compared with the traditional deep neural network model, the accuracy of the algorithm proposed in this paper has increased from 88% to 94.7% in a single learning sample, and the model proposed in this paper does not rely on fine-tuning and controls the actual calculation amount. The handwritten note recognition effect is better, and it has good application prospects in digital music classrooms.