Research On Feature Extraction And Classification Method Of Bird Sound

In recent years,the number of birds in the world has been decreasing.The situation of species diversity is severe.In order to study the reasons that affect the welfare of birds,ecologists need to monitor the activities of birds.At present,the monitoring of bird activities mainly uses the hearing of ecologists and the analysis of the waveform and spectrum of bird sound to identify bird species.This monitoring method is time-consuming and expensive.With the development of wireless acoustic sensors and machine monitoring technologies,ecologists can remotely collect bird sounds by deploying wireless acoustic sensor networks instead of on-site collection.Due to the influence of environmental noise and imbalanced data distribution,North American warblers perform poorly in sound recognition.To further improve the accuracy of bird sound monitoring,this paper takes the sound of 43 species of North American forest warblers(Parrotidae)as the research object.Based on their vocal characteristics and spectral structures,appropriate feature extraction and classification methods are studied to pvovide supplementary information for migratory bird migration research.The main contents of this paper are as follows:Firstly,according to the acquisition process and acoustic characteristics of the North American forest warbler,an appropriate preprocessing method is selected to unify the length and amplitude of the bird sound.The traditional voice recognition algorithm,based on manual features and deep learning,is analyzed.After the manual features,described by acoustics and image are generated.The machine learning algorithms,including K nearest neighbor and support vector machine,are used to determine the species information in the bird sound.In the deep learning,VGGNet,Mobile Net and Efficient Net are used to determine the species information in the bird sound.The experimental results of the traditional bird sound classification are analyzed as the baseline of this research.Secondly,in view of the characteristics of short duration and high frequency of North American wood warblers,one-dimensional local binary pattern and one-dimensional local ternary pattern are combined to divide bird sound into small time segments.It is used to describe the sound by integrating the texture characteristics in a short time.In order to generate the deep features of bird sound,the migratory bird sound is decomposed by 9 layers of wavelet based on the "sym4" wavelet.The high frequency coefficients are removed.The low frequency coefficients of each layer and the texture features of the waveform are integrated.Meanwhile,the DCT and DFT are combined to convert migratory bird sound into frequency domain and to generate deep texture features of frequency coefficients.Aiming at the large dimension and high redundancy of texture features,an improved feature selection algorithm is studied,which selects the optimal features and uses K-nearest neighbors and support vector machines for classification.To reduce the influence of imbalanced learning,the ensemble learning is used to fuse different features.The balanced accuracy of 88.70% is obtained.Thirdly,aiming at the defect that texture features are difficult to further improve the accuracy of bird sound.Multi-dimensional neural networks are introduced as a feature extractor to generate deep features of migratory bird sound,including 1D CNN-LSTM,2D VGG-Style and 3D Dense Net121.Models of different dimensions can extract different frequency dynamic information of bird sound.To reduce overfitting,Mix up data augmentation is used to generate virtual data to help model training.The deep features of different models are fused and classified by k-nearest neighbor and support vector machine.Experimental results show that the model can achieve a balanced accuracy of 93.89%,which can effectively improve the accuracy of migratory bird sound.Forthly,in order to facilitate ecologists to use the algorithm to analyze migratory bird sound,this paper uses the Python language to develop a bird sound recognition system based on the Py Qt5 function library and tools such as Qtdesigner and Qtui.The human-computer interaction is carried out by means of buttons.Functions,such as waveform analysis,spectrum analysis,feature extraction and species identification of migratory bird sound,are realized.