Research Of Music Genre Classification Based On Convolutional Neural Network

With the rapid development of Internet technology and digital multimedia technology in recent years,the number of artworks like literature,photography and music has exploded.More and more people have participated in the creation,dissemination and appreciation of artworks through the Internet.For one of these types like music,various types of online music services have become the main channel for people to listen to music everyday.People's personal preferences for music genres have also prompted service providers to provide more accurate classification results for genres.The traditional way of understanding the music and manually classifying it by professionals is overwhelming when facing the massive music works,it's inevitable to use computer programs to automatically classify the music works.In the task of music genre classification,many classical machine learning methods have achieved state-of-the-art results on standard data sets,but these methods use a lot of hand-crafted features designed by domain experts,the barrier is high for people who are not domain experts.Besides,some of these features lack versatility and cannot be well migrated to other domains.With the extensive use of deep learning models in other fields,there have been more and more methods that are using deep learning models in the task of music genre classification.The recent deep learning methods are still insufficient in terms of accuracy,complexity and training.This paper proposes a new convolutional neural network architecture—Densely Connected Inception Neural Network(DCINN),to solve existing problems.This paper draws ideas from two important convolutional neural network architectures in the field of computer vision,the DenseNet and the Inception module,and proposes a new convolutional neural network structure named DenseInception module.Based on this module,the DCINN is proposed as a new convolutional neural network architecture to be used in the task of music genre classification.The DCINN improves the information flow between the input and the output of the network using the strategy called dense connectivity,which is also used by DenseNet.Besides,The DCINN uses Inception-like module to select the effective convolution kernel size automatically,which improves the representation skills and reduces the burden for model designers.The DCINN adopts the prediction method based on audio clips,which pays more attention to learn the local effective features of audio data,and gives the model the ability to classify audio samples with different lengths.This method makes the model support for classification tasks in different scenarios.In this paper,several experiments have been carried out on the GTZAN dataset and the ISMIR2004 dataset.The results show that the DCINN exhibits powerful capabilities in feature extraction and classification.The 10-fold cross-validation accuracy results on the GTZAN dataset and the ISMIR2004 dataset achieve 88.7% and 87.68%respectively,surpassing the leading models using similar methods,second only to small number of classical machine learning models using hand-crafted features designed by domain experts.After pre-training with the MSD dataset,the performance of DCINN is further improved.The accuracies for DCINN on the GTZAN dataset and the ISMIR2004 dataset increase to 91.0% and89.91% respectively,which proves the validity and advancement of the proposed model.