FCLS: A lightweight model based on sound feature fusion for bird classification

By automatically detecting and classifying bird sounds, relevant researchers can learn ecological information about bird distribution, migration patterns, and habitat preferences. This paper proposed a bird sound classification method based on a lightweight model Feature fusion in bird sound Convolutional Neural Network (BS-CNN) and Long Short Term Memory(LSTM) with Support Vector Machine (SVM), abbreviated as FCLS. Among them, the BS-CNN integrated a self-attention module. Firstly, spectral subtraction was used to denoise the original sound data. To address the problem that the traditional double-threshold method struggled to extract segmented bird sound samples in their entirety, a triple-threshold method was designed to automatically extract segmented bird sounds from complex sound data. Then the sound features, Mel Frequency Cepstrum Coefficient (MFCC) and Gammatone Frequency Cepstrum Coefficient (GFCC), were extracted and fused as the fusion feature Mel-Gammatone Frequency Cepstral Coefficients (MGCC). Finally, the MGCC was input into FCLS to obtain the classification result. There were a total of 10 bird species and they were segmented into 10,480 audio samples by the triple-threshold method. To verify the efficiency of the proposed model, FCLS was used to compare with 12 deep learning models and 4 traditional machine learning models based on the self-built bird sound dataset. The results verified the superiority of the FCLS.