One-step progressive representation transfer learning for bird sound classification

Abstract

Most bird sound classification methods are based on supervised learning, which requires a large amount of labeled data for neural network models to establish the mapping relationship between bird sounds and labels. With the advent of self-supervised learning methods, neural network models can learn more general representations from unlabeled raw data and perform well on downstream tasks compared with fully supervised learning. In this study, we find that the self-supervised, pretrained representations can be transferred to the bird sound classification task. Furthermore, the model trained through self-supervised representation learning achieves a higher recognition rate after fine-tuning compared with the model based on supervised classification learning. However, this improvement in performance comes at the cost of the pretraining time amount. Considering these problems, we propose a one-step progressive representation transfer learning method for bird sound classification. This method integrates self-supervised representation and supervised classification learning into a two-branch network and uses a time-dependent loss weight transfer strategy to transfer bird sound, and self-supervised representation learning to bird sound classification learning. By using this method, the model emphasizes self-supervised representation learning during the early stages of training to maximize the similarities among bird sounds across different data augmentation versions. By utilizing the loss weight transfer strategy, the self-supervised bird sound representation learning transitions to supervised bird sound classification learning, thus enabling the model to acquire specific bird sound classification capabilities in the later stages of training. Experiments using the dataset Birdsdata demonstrate that our method outperforms both supervised classification and self-supervised classification learning methods.