Research On Chinese Opera Classification Based On Deep Learning

Chinese opera has a long history and contains the essence of traditional Chinese culture.However,in recent years,traditional Chinese opera has faced serious development problems that many local operas are on the verge of extinction.Using computer technology to study the automatic classification algorithm of Chinese opera is beneficial to the management and protection of Chinese opera.In recent years,great progress has been made in music classification methods based on deep learning.At present,there are still few studies based on Chinese opera classification.It is possible to use the deep learning-based music classification method to classify Chinese operas.However,there are still issues such as feature extraction without considering the rhythm of the Chinese opera,modeling without considering the context dependence between the segments of the Chinese opera,and huge model parameters.Aiming at these problems,we propose a Chinese opera classification algorithm based on deep learning.Its main research contents and innovations are as follows:(1)In terms of feature extraction,we propose a time-frequency feature extraction method based on variable Q transform and filter banks technology.We first build a high-quality data set by collecting Chinese opera repertoires that contain eight Chinese operas.Then we use the variable Q-transform to transform the spectrum of the Chinese opera signal.Through the filter banks designed based on the Chinese opera rhythm,the energy distribution of each frame in the Chinese opera segment signal at the scale frequency is acquired and spliced,thereby extracting the time-frequency characteristics.The method overcome the problem of insufficient time resolution in the low frequency band of the mainstream method,and also consider the influence of the pentatonic scale of the Chinese opera and the "cavity" factor,which is more in line with the rhythm of the Chinese opera.(2)In terms of model design,we propose a Chinese opera classification method based on a multi-layer cascaded neural network.It contains two-level models.The first-level model is composed of a one-dimensional residual network and a multi-scale feature extractor to learn the time-frequency characteristics of the Chinese opera segments.The second-level model is composed of a bidirectional recurrent neural network,which learns the context dependency relationship between each segment in the Chinese opera.Furthermore,we introduce an attention mechanism to learn the importance between each segment.This method not only considers the internal characteristics of the segments but also the context dependence between the segments.Experiments show that the overall accuracy of the proposed method reaches 96.5%,which surpasses the classification algorithm based on convolutional neural network with voting and the classification algorithm based on traditional machine learning.(3)In terms of model optimization,we propose a more lightweight and parallel multi-layer cascaded neural network which is deployed on the embedded side.Because the parameters of the multi-layer cascaded neural network proposed in this thesis are large and the calculation efficiency is low,here we further use the deep separable convolution and Transformer structure to achieve lightweight and parallelization of the first-level model and second-level model of the multi-layer cascaded neural network,respectively.The optimized model is deployed on the embedded platform Nvidia Jetson? TX1,and the Chinese opera classification system is implemented on the embedded platform.Experiments show that this method can reduce the parameters of the model,speed up the calculation,and realize the classification of opera on the embedded side,while ensuring the accuracy of the model.