Research On Music Classification Methods Based On Deep Learning And Transfer Learning

As a key component of music information retrieval,music classification is an important method to deal with massive music information.In recent years,with the increasing demand for music content retrieval,the research of music classification mainly focuses on the methods of music genre and emotion classification.However,manual annotation which the traditional music classification method uses is costly and not scalable.As a result,automating classification is the main research direction.At present,most of the automating methods use manual extraction of acoustic digital features as the input of traditional classifiers,but it is a great challenge to the realization of engineering that such a way relies on heuristic knowledge.This paper introduces deep learning to solve this problem.Deep learning technique has received intense attention owing to its great success in image recognition.Powerful feature extraction ability can replace complex feature engineering.In this paper,we conduct a comprehensive study on music audio classification with improved convolutional neural networks(CNNs).For the classification of music genre,this paper uses the Spectrogram as the input of convolution neural network,and proposes a one-dimensional convolution neural network model(1D-CNN),which takes the sound axis as the width,the frequency axis as the channel,taking into account the audio timing characteristics and the local perception characteristics.In this paper,we exploit the advanced DenseNet of CNN architectures to boost performance of music audio classification,and propose to use data slices for network training.The process is using the trained network as feature extractor,parallel slices through the network for average collection of information to form music feature vector,finally using support vector machine(SVM)for classification.In the part of data processing,music-specific data augmentation is realized with the time overlapping and pitch shifting of spectrograms.This paper introduces the proposed model(Dense+SVM)in detail,and proves that Dense+SVM can effectively improve the classification accuracy through experiments we designed.To address the shortage of labelled audio data in music emotion classification and annotation,this paper proposes the use of transfer learning for classification,mainly introduces two transfer methods.One way is to transfer the convolution part of the model realizing the knowledge sharing of genre and emotion,then to use the extracted features for SVM classification.The second method is concatenated feature transfer learning,which combines the low-level information with the high-level information by concatenating the convolution features,and proposes to use PCA to descend dimension by selecting the feature principal components.Experiments showed that the proposed the concatenated feature and dimension reduction method can effectively improve the accuracy of emotion classification.