Research On Bird Song Recognition Algorithm In Natural Scene

As an important part of the ecosystem,the monitoring of bird activity and distribution provides an important basis for understanding changes in biodiversity and climate change in an area.Bird song is an important feature for distinguishing birds and is currently one of the commonly used methods for bird species identification.Bird monitoring through bird song has the advantages of being efficient,stable and wide-ranging,and has great application value.The design of a robust and noise-resistant bird song identification method is important for understanding bird biodiversity in complex natural scenes where there is often a large amount of environmental background noise.For these reasons,this paper investigates bird song recognition algorithms in natural scenes.The research is divided into three main parts: recognition methods based on acoustic features and machine learning,recognition methods based on deep learning networks and the development of a prototype system for bird song recognition.(1)To address the problems of single extracted features and low classification accuracy in bird sound recognition algorithms,a joint hybrid feature selection and grey wolf algorithm based kernel limit learning machine bird sound recognition method is proposed.First,the acoustic feature set Com Par E is extracted from the bird sound data;second,the Fscore of each feature is calculated and ranked;then,the generalized sequential forward floating search is used as the search strategy,and the kernel limit learning machine and the ten-fold cross-validation strategy are used to optimally select the features to obtain a subset of features suitable for bird sound recognition;finally,the optimal kernel limit learning machine is selected by the grey wolf algorithm Finally,the optimal kernel limit learning machine parameters are selected by the grey wolf algorithm to recognize bird sounds.The experimental results show that the method can effectively improve the recognition accuracy of the Com Par E feature set in the field of bird sound recognition.(2)In order to enhance the feature learning ability and improve the recognition accuracy of bird song signals in natural noise environment,a bird sound recognition method based on deep residual systolic network and dilation convolution is proposed.Firstly,the Log Mel feature set is extracted from the log Mel signal and its first-order and second-order difference coefficients as the input of the network model.Finally,the network learns long-term dependencies from the learned local features.The experimental results show that this model outperforms existing models in noisy environments.(3)Based on the deep residual systolic network and the expanded convolutional bird sound recognition method,a more efficient bird sound recognition method is designed by combining the lightweight model Mobile Net V3.Firstly,the SE module in the shallow network of the V3 model is replaced by a deep residual unit so as to achieve noise suppression,and then the Mobile Net V3 network is trimmed to further reduce the number of model parameters.Finally,the trained model is applied to a prototype bird song recognition system using the Beijing Bird Database to monitor 19 bird species in the database.