Research On Dolphin Species Identification Based On Machine Learning

Dolphin is a cetacean mammal with a developed sonar system.The traditional recognition of dolphin sound signal needs to extract dolphin sound signal manually from a large amount of dolphin sound data,which is time-consuming and labor-intensive.Because the dolphin echolocation signal(Click)is a single pulse signal and the echolocation signal of different dolphins is different,it can be used to identify dolphin species.In addition,the traditional identification of dolphin species requires visual observation,so it is difficult to identify and track dolphin species in real time in the dolphin detection network.In recent years,there are more and more researches on machine learning methods,which can be used for data classification and prediction.Therefore,this research takes dolphin echolocation signal and dolphin image as recognition objects respectively,and uses machine learning method for automatic recognition of dolphin species.In this paper,five machine learning methods are used to identify dolphin species based on dolphin echolocation signals.In this experiment,the Sousachinensis in Xiamen bay and Leizhou bay were set as the first group to identify dolphins of the same species.The Sousachinensis and bottlenose dolphin in Xiamen bay were set as the second group to identify different dolphin species.The time-frequency diagram of dolphin sound signal was firstly calculated,and the approximate position of Click signal was determined by adaptive threshold and connected domain analysis.Then the Teager-kaiser energy operator(TKEO)and Gabor filter are used to precisely locate the Click signal.Then the Click signal is extracted by cepstrum method.Finally,five machine learning methods were used to identify dolphin species.The experimental results showed that the average recognition accuracy of the first group and the second group was above 99%and 96%,respectively.Then,a method based on convolutional neural network is proposed to identify the dolphin species in the first group of the same species,and the optimal selection of parameters in the convolutional neural network is discussed.When the convolution kernel size of 1 or 3 is not suitable for dolphin Click signal classification,the classification accuracy is low.When the convolution kernel size classification using 9 or 11 is effective,the recognition accuracy of the test can reach 99.75%.For the second group of different dolphin species recognition,a method based on Xgboost was proposed to improve the recognition accuracy of different dolphin species,the result show that the recognition accuracy of this method can be improved to more than 99%.At last,the image of dolphin is taken as the object of recognition and the method of convolutional neural network is used to identify dolphin species.In the experiment,firstly,the crawler technology was used to obtain dolphin images from the Internet,and then the SSD model was used to detect the dolphin images.Then,by comparing the recognition effect of different proportions of training sets,the recognition effect before and after the expansion of data sets,and the recognition effect of three models(AlexNet?VGGNet?ResNet),respectively.Finally,the cutting proportion of the training set was 90%,the data set was enlarged by image mirroring,and the ResNet model was used for dolphin image recognition.The training accuracy of the model is 97.94%and the testing accuracy is 93.06%.This research shows that machine learning methods can analyze dolphin sound signals and dolphin images,and automatically identify and detect dolphins in the sanctuary,which can better study and protect dolphins.