Research On Recognition Algorithm Of Digital Piano Score Difficulty Level

Facing a vast number of digital piano music resources on the Internet,whether music amateurs or professional learners all find it difficult to search suitable music scores for their own learning difficulty level.They will get lost in the massive music database easily.In music recognition field,many studies have focused on music genre recognition,music emotion classification and so on.But there is less research on the recognition of the difficulty level of score.In order to solve the problems such as it is difficult to assign difficulty level labels for mass music by human,and the current classification method is highly dependent on the individual and the existing algorithms' recognition accuracy is not good enough.In this thesis,we design a kind of improved automatic algorithm for the difficulty level recognition of the piano score.Firstly,in order to find more difficult related features and enrich the feature space,after consulting a number of professional piano teachers and analyzing the difficulty classification criteria of music website,some new features are proposed and then verified by contrast experiments and ReliefF weight algorithm.Then using the theory of metric learning to make full use of the prior knowledge of training data and find a new distance metric to improve the k nearest neighbors(KNN)algorithm's performance,we propose a feature projection KNN algorithm----P-KNN to realize music scores' difficulty level recognition.In addition,the thesis further explores the performance of ML-SVM algorithm which is established by using metric learning to improve Gauss radial basis kernel function,in music scores' difficulty level recognition.Simulation experiments are done in two score data sets with different characteristics.The proposed P-KNN and ML-SVM algorithms are compared with other popularly used classification methods.The results show that the proposed algorithms' recognition accuracies are better,reaching 83.5% and 84.67%,respectively.Classification performances are effectively improved than the original KNN or original SVM based on Gauss radial basis kernel function.