Multi-label bird species classification using Haar wavelet- based residual convolutional neural network

Automatic bird vocalization analysis is advancing research in ecology and conservation. In recent years, numerous studies have employed deep learning models to categorize bird calls. This study examined the efficacy of Haar Wavelet Residual Convolutional Neural Network (WRCNN) for multi-label bird species classification. Initially, Haar wavelet transforms were applied to the mel spectrograms of bird call recordings. These transformed spectrograms were subsequently input into the WRCNN for multi-scale spectral analysis. The model obtained a macro-average F1-score of 0.60, showcasing its potential in multi-label tasks and exhibiting notable improvements over baseline methods. Experiments were conducted utilizing the Xeno-Canto bird sound database.