Whale Calls Classification Based On Transfer Learning And Improved Structural Depth Neural Network

How to quickly identify the passive acoustic signals of whales in a long range has always been one of the problems that researchers focus on solving.This paper takes the typical whale calls as the standard target samples.Convolution neural network and generative confrontation network are used to recognize the calls of humpback and blue whales.From the feature extraction,network design and other aspects of the algorithm design and implementation details.The main research contents are as follows:(1)First,three feature extraction methods are compared: Mel Frequency Cepstrum Coefficient,triangular filter banks and spectrogram.The results of three feature extraction methods for whale call samples and their advantages and disadvantages in neural network are analyzed.(2)Second,this paper summarizes the convolution neural network,explains the advantages of convolution neural network for classification task,tries to analyze the effect of deep convolution neural network on whale call classification,analyzes the influence of network depth on classification accuracy,analyzes and compares the design logic,advantages and disadvantages of each network structure.The training details of convolution network structure,the principle of neural network optimization and the use of various optimizers are described.In addition,this paper uses transfer learning to solve the problem of insufficient target whale call samples,and verifies the effectiveness of transfer learning through experiments.(3)At last,in this paper,we introduce Generative Adversarial Networks to do sample enhancement from two aspects.First,the use of DCGAN combined with CGAN to expand the number of samples can effectively avoid over fitting and improve the classification accuracy.Secondly,Cycle GAN is used to enhance the features of the samples to reduce the influence of underwater noise on classification accuracy.In order to improve the robustness of the classifier,this paper presents the classification results in parallel with the deep convolution network model by using the discriminative model of Generative Adversarial Networks.Experiments show that deep CNN and GAN working together can bring better classification effect by using spectrogram as feature extraction method,and the structure improvement brought by the fusion of discriminative model and convolutional neural network increases the stability of classification results.