Research On Grading Of Lute Plate Based On Sound Feature And Convolution Neural Network

The acoustic quality of the Chinese national musical instrument resonance component is the key factor to evaluate the quality grade of the instrument,and the wood is an important material for making the resonance component of the national musical instruments.Therefore,the research on the acoustic quality of the wood has important significance for the evaluation of musical instrument quality grade.At present,in the production process of musical instruments,the choice of instrument resonance components is mostly judged by the experience of technicians.This method has the disadvantages of long time,high labor cost and insufficient judgment basis.In order to solve this problem,this paper selects a representative national musical instrument lute as the object,and studies the acoustic vibration performance of the wood used to make the resonance component(resonance board,resonance box),and extracts the characteristic parameters which can effectively characterize the acoustic performance of the lute resonance wood component,then establish a corresponding acoustic classification model.At present,sound recognition and deep learning technology have been developed rapidly and made great progress in the field of sound and image.So the focus of this paper is to apply deep learning and sound recognition technology to the field of lute material selection in order to realize intelligent material selection.This research will play a positive role in promoting the improvement of the technical level of lute selection in China.Firstly,aiming at the problem of insufficient distinguishing ability of original sound features,this paper improves the method of sound spectrum image feature extraction of wood sound signal.Firstly,the wood sound signal is preprocessed,then transformed into the image feature of sound spectrogram.In order to improve the feature expression and detail information of the sound spectrogram,the improved CLAHE method is used to process the sound spectrogram.The spectrogram is converted into the HSI space by the color space transformation method,and the I component is enhanced by the improved CLAHE and bilinear interpolation method to obtain the enhanced spectrogram.Finally,in order to verify the validity of the enhanced spectrogram features,the Mel frequency cepstral coefficient(MFCC)and differential features,original spectrogram features and enhanced spectrogram features were compared as input to the classifier respectively.The enhanced spectrogram features extracted have a better classification effect on distinguishing wood sound signals.Secondly,for the problem of low recognition accuracy of classical classification methods,this paper compares the classification performance of classical algorithms and deep learning models.Among them,the classical algorithm uses Gaussian mixture model and support vector machine as the baseline system,and the deep learning model selects the convolutional neural network(CNN)structure.Aiming at the difficulty in selecting the parameters of CNN model,a method combining grid search and dynamic learning rate is proposed to solve it.Considering the problem that the depth of the network layer has a great impact on the results,three different depths of CNN structure are set.The network structure with the best depth is improved to obtain SMCNN.Finally,in order to further optimize the SMCNN model,the advantages and feasibility of using the ELM as the SMCNN classifier are analyzed,and the hybrid model of SMCNN-ELM is obtained.The experimental results show that the ELM algorithm is introduced into the convolutional neural network.Compared with the simple use of convolution neural network model,this method can effectively improve the recognition accuracy,and can solve the problem of too long training time.In summary,the proposed scheme uses wood sound signal as the entry point,improves and optimizes the two aspects of sound feature extraction and classification algorithm,which provides a reliable idea for the field of lute selection and research on grading of lute plate.