Enhanced Voiceprint Recognition For Birds By Information Fusion

Birds play an essential role in ecosystems,effectively monitor and identify their distribution are of great significance for the protection and assessment of ecosystems.The survival and development of rare wild birds maintain the balance and stability of the ecosystem and it's crucial indicators for evaluating the quality of the ecological environment.At present,the classification and identification of bird species is mainly through birdsong data,but the birdsong data in a natural complex acoustic environment usually has problems such as complex environmental noise,unbalanced song data and incomplete biological spectrum information.In recent years,the brilliance of deep learning technology has enabled the identification of bird species with large-scale dataset.Seeking an efficient and robust bird species identification method is of great significance for protecting bird species diversity and evaluating ecosystems.Based on the above reasons,under the background of bird species recognition,combined with the multi-dimensional birdsong characteristics,the research is suitable for the recognition method of bird voiceprint in natural acoustic environment.The research work is divided into three parts: birdsong feature enhancement method based on image processing,birdsong spectrogram feature set based on feature fusion and multi-channel bird voiceprint recognition model based on decision level fusion.(1)In the pre-processing stage of birdsong,a processing strategy is proposed for the characteristics of birdsong which from noise reduction of birdsong spectrum,song signal cutting and feature extraction of birdsong spectrum.And through the combined use of a series of image processing methods,the three functions of bird noise spectrum signal-to-noise separation,specific data enhancement and visual perception enhancement are realized,and the interpretability of bird song data is enhanced.(2)In the birdsong feature extraction part,five voiceprint features of birdsong are selected,and three aggregate features with multi-scale time-frequency characteristics are generated based on feature fusion.Through the fusion and complementation of multiple features,the sensitivity to environmental noise when a single feature is classified is reduced,which helps to improve the spatial coverage and enhance the robustness of the overlapping of the chirping features.By fusing multi-level features,time-frequency domain features,local and global features,intelligent birdsong spectrogram information processing is realized.(3)In terms of bird voiceprint recognition model,a multi-channel convolutional neural network model based on decision-level fusion is proposed.The model combines the time-frequency characteristics of multi-scale birdsong sounds,uses spatial information to compensate the low signal-to-noise ratio song information,and maintains accuracy while improving sensitivity to birdsong features.The experimental results show that the performance of the merged neural network with decision-level fusion function is better than that of a single deep architecture.The classification accuracy of this structure can not only exceed LTCNet,MTCNet,and GTCNet,but also outperform existing high-performance CNN models.Through the above three steps,information fusion has been implemented to enhance bird voiceprint recognition which forming an intelligent birdsong spectrogram data information processing and a robust multi-channel bird voiceprint recognition model.Finally,combined with the list of rare wild protected birds in Hubei,52 key protected birds were identified.The experimental results show that the birdsong feature enhancement method based on image processing effectively improves the problems of high background noise,unobtrusive song feature and incomplete bio-spectrum information in the recognition process of birdsong audio signals;The aggregated features based on data-level fusion are more robust than single voiceprint features;the bird-based voiceprint recognition system based on information fusion is superior to the typical high-performance convolutional neural network architecture in classification accuracy.