Small sample optimized bird sound recognition network based on bridging transformer
In view of the uneven distribution of bird sound audio data collected by actual bird monitoring,the neural network training is not sufficient,and the classification recognition test accuracy is low,a bridging Transformer neural network model is designed.The network first uses the original birdsong audio signal to generate the short-time Fourier transform spectrogram as the input feature,and then inputs the spectrogram into the Transformer network composed of the attention module and the convolution module to complete the information interaction of the global and local features in the spectrogram.Finally,the single-layer Transformer encoder is used to optimize the loss in each batch of samples to obtain the final classification result.Small sample experiments were carried out on Birdsdata and xeno-canto bird sound datasets,and the average accuracy rates of 91.34％and 82.63％were obtained,respectively.Comparative experiments were carried out with other bird sound recognition networks to verify the effectiveness of the network.
万方数据
维普
