Research On Music Genre Classification And Cover Song Identification Based On Deep Convolutional Networks

In recent years,the copyright industry of digital music products has developed rapidly with the popularization of Internet technology.However,the copyright classification management of digital music products is chaotic,and the problems of lack of cover infringement protection are still difficult to solve.Based on the deep convolutional network,this paper combines the music genre classification method and the cover song similarity identification method,and applies it to music copyright classification management and cover infringement protection,strengthens music classification management,optimizes cover identification efficiency,and promotes the construction of digital music product copyright platform.The main works of the thesis are:Music feature generation network based on genre label constraint.Existing music genre datasets have the problem of unbalanced data volume of each genre.The paper uses the Auto-Encoder for rare types,combined with existing music genre labels to impose genre constraints on the music features generated by the feature decoder,generated music features of the same genre as the input music features but with different content,thereby reducing the impact of data imbalance on genres.Information summary enhancement and interactive fusion of music genre classification model.The model enhances the ability to summarize information in the time domain by expanding the scope of the receptive field,so as to better capture genre feature information.In order to preserve all genre information,adaptive average pooling is used,and a multi-channel 1*1 convolution kernel is generated as feature excitation,which realizes cross-channel fusion of information,enhances the nonlinearity and improves the genre classification ability of the model.Feature fusion and clustering cover song similarity identification model.There are many changes in the music elements of cover songs,and it is difficult to identify.This paper applies the autoencoder-based clustering method to cover song identification for the first time,proposes a music feature clustering module,generates music feature clustering labels,and enriches the types of labels.The fusion feature extraction module is proposed,and the dual input features are combined along the channel dimension,which realizes the whole process of feature fusion learning and improves the model identification performance.The feature clustering labels and music version labels are combined through a channel separation decision module to jointly optimize the binary classification objectives of cover song infringement and non-cover song infringement.