Whale Call Recognition Based On Hybrid Model

This paper takes actual marine whale call signals as the research object,extracts the auditory perception features of five typical whale calls under the background of marine noise,and recognizes them by combining machine learning classifiers and statistical distribution models.Taking into account the shortcomings of the low recognition rate of machine learning,two typical deep learning basic models are introduced,CNN and LSTM.Correspondingly improved its performance,designed a variety of hybrid models,and gradually improved the recognition rate and stability of the network.The background,purpose and significance of the related research on whale call detection and recognition are introduced,the domestic and foreign current situation of whale call research is analyzed,and the development process of whale call recognition is explained along the two lines of feature extraction method and classification decision model.Then,the theoretical basis and performance advantages of deep learning are discussed separately.In this paper,auditory perception features are introduced and the feature extraction process is optimized.Then,the DBI indicator is used to evaluate the distinguishability of various features of whale calls.In addition,we use different machine learning classifiers to identify based on different features.In view of the different "sensitivities" of different features to different types of classification models,ensemble learning methods are applied.On this basis,a heterogeneous integration improvement method that uses multiple features to increase the "diversity" of the base learner is proposed,and the recognition effect is greatly improved as a result.Hidden Markov model has outstanding performance in the field of speech recognition and natural language processing,but few studies have combined it with mammalian call recognition.Therefore,this paper builds a hidden Markov mixture model based on statistical distribution,uses Gaussian mixture distribution to describe the observation matrix,and obtains a performance much higher than that of a single machine learning classification model,and demonstrates the unique position of this method in the field of audio recognition.However,the performance of the two in convergence and stability is obviously not satisfactory,so on this basis,a CLSTM hybrid model is designed,which combines the powerful spatial feature extraction capabilities of CNN and the advantages of LSTM in timing signal processing,and has good recognition results.However,the stability of the model is still not ideal enough.At the end of the article,an improved multi-channel input CLSTM hybrid model is proposed,which not only obtains a higher recognition effect,but also optimizes the robustness of the model.Using the SVM classifier to replace the Softmax layer of the neural network,the optimal recognition rate of the single model is 98.24%.Using the idea of permutation and combination,the above four models were heterogeneously integrated,and the highest recognition rate of the full text was 98.69%.