Birds Species Identification Algorithm For Two-modal Feature Fusion Based On Deep Learning

Diversity survey of bird species is one of the major contents to study the structure and function of bird community.Bird sound is one crucial biological characteristics of birds as it presents ideal recognition,which has been widely used in bird species classification and behavioral research.In recent years,the newly proposed method of using automatic recording equipment and recognition software to realize bird survey through vocal recognition has great application prospects thanks to various merits such as high efficiency,non-injury,low interference and wide-ranging monitoring.To realize the needs of bird statistical analysis,this thesis aims to conduct study on the recognition of random disturbance noise in the natural environment based on the recognition of random disturbance noise,the difficulty in manually extracting the differences of bird songs and the low recognition efficiency of recognition models.The research on sound processing,species identification and application is carried out.The main contribution of this thesis is presented as follows.1.A method for bird species identification based on migration learning was proposed.Regarding the high data requirement and the limited data source of birds,the deep classification technology proposes a species classification method based on feature and model migration learning.After the noise removing of the audio signal in the natural environment using the single channel blind source separation algorithm,The time-frequency analysis method was used to obtain the speech map of the denoised audio signal.After the expansion of data enhancement technology.the source image database was built.The Convolution Neural Network(CNN)was used to automatically extract the high-level features of the language map and complete the classification,providing necessary research basis for other bird project applications.2.A classification algorithm for bird species based on two-modal feature fusion was proposed.This thesis combined the structure of convolutional neural network and Long Short-Term Memory(LSTM)to establish a more generalized Convolution Long Short-Term Memory(CLSTM)network model.The optimized network structure includes input pre-processing,learning rate,prevention of over-fitting,gradient disappearance.Convolutional network extracted features and long-short-time memory structure extracted sound sequence characteristics were fused and adaptive species identification was realized based on tweets or songs.The optimization of the sound classification algorithm was realized and the recognition effect is improved.3.A mobile embedded bird identification system based on MVC(Model View Controller)architecture was designed.This thesis used Eclipse developing tools to develop a bird recognition software for the practical collection of sound signals.The client APP integrates ftunctions such as data module,view image module and control module.The human-computer interaction interface developed in this thesis presents high recognition performance and contributes to the automatic recognition task of individual birds based on Android,which greatly promotes the automation of bird resource investigation and monitoring.In order to verify the effectiveness of the proposed algorithm,18 bird audio samples collected by the bird open data set Xeno-canto and the Songshan National Nature Reserve in Beijing were used to evaluate the algorithm through application experiments comparison with traditional algorithms and models.The experimental results showed that the proposed method presents a great improvement in the performance evaluation indicators such as correct rate,accuracy,recall rate and F1 value.The automatic classification and statistics of bird species based on the sounds are realized.The method can also be applied to the automatic statistics of the quantity of other birds,which has important theoretical significance and engineering application value.