Optical Music Recognition Algorithm Combining Multi-scale Residual Convolutional Neural Network And Simple Recurrent Units

As a key technology of symbolizing music score,optical music recognition is beneficial to the storage and editing of music files,and has important application value in music information retrieval and computer-aided teaching.Optical music recognition algorithm based on general framework faces the problem of complex processing steps and low precision.The algorithm based on the deep learning efficiently simplifies the general framework.However,the recognition accuracy needs to be further improved,and there is more error when identifying difficult notes.The training of models takes more time.Thus,this thesis proposes an optical music recognition method based on improved convolution recurrent neural network.The proposed algorithm of optical music recognition in the thesis is composed of four parts.Firstly,some music images are enhanced in the data set to expand the music images data and improve the robustness of the training model.Next,the residual convolutional neural network is used to extract notes features in the music images to deal with the degradation of the model.Then,multi-scale features are fused into the same feature map,so as to enhance the representation ability of features and improve the subsequent recognition accuracy.Finally,a network combining bidirectional simple recurrent units with connectionist temporal classification is adopted to identify notes.A large number of calculations are parallelized to accelerate the convergence of training.There is no need for the data set to be strictly aligned with the label,which looses the quality requirement for the data set.By four groups of experiment the validity of the algorithm are evaluated,which include the robustness of the model with the data enhancement,representation ability of the features via residual convolutional neural network,multi-scale fusion effect,and the influence of simple recurrent units on the convergence rate of the model.The experiment result shows that symbol error rate of the improved network model is reduced to 0.3234% on average and the training time of the model is about one third of traditional convolution recurrent neural network,which is optimized in terms of identification accuracy and training time.In addition,the performance of the proposed algorithm,the algorithms mentioned in the existing literature and three kinds of commercial optical music recognition software are compared.The results show that the proposed algorithm is better than the existing algorithms in identification accuracy and training time.Compared with commercial software,the proposed method is more robust,and has better recognition result of the notes such as symbol beams and rest.