A transfer deep residual shrinkage network for bird sound recognition

Bird sound recognition has important applications in bird monitoring and ecological protection. However, in complicated environments, noise and insufficient sample data are the major factors affecting recognition accuracy. We proposed a bird sound recognition method based on a developed transfer deep residual shrinkage network. First, a deep residual shrinkage network with noise resistance was constructed based on the structural characteristics of the residual shrinkage module, multi-scale operations, and the characteristics of bird sound Mel spectrograms. Then, the deep residual shrinkage network was pre-trained using a bird sound dataset, applying an unfreezing finetuning strategy, to mitigate the impact of insufficient training data. A transfer learning alleviated the problem of data scarcity by utilizing pre-trained models, while the deep residual shrinkage network enhanced the performance of the model in a noisy environment by optimizing the network structure. Experimental results showed that this method achieves high recognition accuracy under noise and small data sets. It has advantages over the compared methods and is suitable for ecological monitoring fields such as bird population monitoring. The method has good application prospects.