GraftGIL-Net: A grafting framework on graph inductive learning networks for multi-label bird audio classification

The automatic analysis of bird vocalizations holds significant promise for conservation monitoring, ecology, and vocal behavioural studies. Existing bird call classification methods face difficulties in detecting multiple overlapping bird vocalizations. To address this challenge, we propose GraftGIL-Net, a grafting framework based on graph inductive learning networks for bird call classification. Initially, a graph named Mel-Graph is created from the Mel-spectrogram of the audio file using a trained DeepConvNet model. This graph is then fed to the proposed GraftGIL-Net for classification. The proposed method is evaluated on the Xeno-canto bird sound database and achieves a macro F1-score of 0.95.