Research On Whale Classification Method Based On Improved Mel Spectrum And Feature Selection

The study of marine mammal call signals is of great significance to the protection and biomimetic research field of marine ecological environment,this paper classifies and identifies four typical whale call signals in the context of marine noise,which mainly focuses on two key steps in the pattern recognition task,namely feature extraction and classification decisionmaking,the main content of the paper includes:(1)Design the whale call signal classification and recognition system.In the feature extraction link,the whale call signal is converted into a spectrogram,and the local binary mode features of the sound spectrum image are extracted.For the classification decision-making section,the support vector machines suitable for different data types are introduced,and the applicable scenarios and performance of common kernel functions are analyzed for the nonlinear support vector machine used in this paper,and the grid search method is introduced for parameter optimization,and the final average classification accuracy of the classification system is obtained by combining the K-fold cross-verification method.The effectiveness of the designed classification system is verified by extracting the local binary mode features of the actual marine whale call signal spectrogram and comparing the classification results of the traditional auditory perception features.(2)In view of the wide application of Mel spectrum in the field of speech signal classification and recognition,the concept of Mel frequency is introduced to obtain the Mel spectrum of whale call signal;The contribution degree of each frequency band in the Mel frequency domain to classification identification is studied,and the Mel spectrum is improved accordingly,and the improved Mel spectrum based on the subband contribution is obtained.The local binary mode features of the Mel spectra before and after the improvement were extracted,and the support vector machine was used for classification,and the classification results verified the superiority of the improved Mel spectra in the whale classification and recognition task.(3)The feature selection method is introduced for feature dimensionality reduction,the definition of feature selection and the general steps of feature selection are introduced,and the performance of the feature selection method based on Fisher Score is analyzed.Considering that the feature selection using pure Fisher score ignores the correlation between features,the concept of maximum information coefficient is introduced,and a feature selection method based on the hybrid model of Fisher score and maximum information coefficient is designed.Experimental results show that the designed Fisher Score and maximum information coefficient hybrid model further removes redundant features and improves the classification accuracy.