Study On Bird Intelligent Recognition Based On Sound Feature Fusion And Attention Neural Network

Bird diversity is commonly used as biological indicators to assess the impact of adverse environmental changes on organisms,as well as changes in the status of many different habitats.The monitoring and identification of birds is essential for biological conservation,scientific research and ecosystem management.Compared with vision-based bird identification methods,the audio-based bird identification methods are the preferred direction for bird species classification,because of the advantages of wide application range,no obstructions,minimal human intervention,and automatic acoustic recognition is regarded as an important technology for bird monitoring and protection.However,traditional bird sounds recognition methods have problems such as high cost,long time consuming and low precision,which limit their application and development.To solve the above problems,this paper starts from the bird sounds,and studies the bird intelligent recognition method based on sound features fusion and attention neural network by domestic and foreign research,feature parameter processing,model building,experimental verification and system application development.The main work of this paper is as follows:(1)A method for fusion and optimization of bird song features is proposed.In view of the limited information carried by the single feature,a new linear fusion method is designed to obtain three fusion feature sets(Log-CST,MFCC-CST and Log-MFCC-CST),based on the Log-scaled mel spectrogram(Log-mel),the Mel-frequency Cepstral Coefficients(MFCC),the Chroma,the Spectral contrast and the Tonnetzs.The verification experiments on the two benchmark models(Dense Net and Efficient Net)are performed to select the optimal fusion feature sets,thus providing the data basis for subsequent applications.In this case,the accuracy of the optimal fusion feature set(Log-CST)is 88.6%for the Dense Net model and 88.2% for the Efficient Net.(2)A bird recognition algorithm based on attention residual network is proposed.This thesis establishes a more applicable attentional residual networks(AMRes Net)based on the attentional mechanism and residual networks.The performance of the model is evaluated by the accuracy,the confusion matrix and the ROC curve.In the test phase,ablation experiments of different amounts of attention layers are used to gradually improve the accuracy of the model,and the finally built model provides a model basis for algorithm optimization.The experimental results show that the AMRes Net model achieves the best classification accuracy(92.6%),compared with the eight commonly used classification models.(3)A bird recognition method based on improved semi-supervised learning strategy is proposed.This paper studies the development of Semi-Supervised Learning(SSL)strategy,proposes an improved semi-supervised learning strategy(ARMatch),to optimize the bird recognition model(AMRes Net).The performance of the model is effectively optimized on the weak label data sets by the design of several key steps,and the accuracy,precision,recall and F1 score are used to verify the effectiveness of this method.Compared with the supervised learning method and the two semi-supervised learning methods,the ARMatch method achieves the best recognition results on the two sound data sets,namely 0.47±0.03 or 0.87±0.03(the number of labeled samples is 100 or 1000),0.50±0.01 or 0.92±0.04(the number of labeled samples is 150 or 1500),0.52±0.02 or 0.95±0.13(the number of labeled samples is 200 or 2000).(4)A bird sound intelligent recognition system based on HTML5 and Flask is designed.Based on the bird recognition model,this thesis develope a set of response modes of integrated data uploading,front-to-back interaction,model embedding,and interface display by using Flask framework,as well as complete the functional layout of the client interface with HTML5 technology.