Bird Song Recognition Based on Deep Transfer Learning with XGBoost

Due to the influence of many factors such as geography, environment, body size and age, the performance of bird song classification depends heavily on the effectiveness of extracting differentiated features from the audio signal and building efficient audio classifier. In this paper, we first calculate delta and delta-delta for log-Mel spectrogram (log-Mels) to form 3-dimensional (3D) time-frequency spectrogram based on log-Mels not only preserves species differentiation but also reduces the influence of bird song-independent factors in the actual scene. Second, we propose a novel audio classifier based on deep transfer learning combining VGG16 with XGBoost. Specifically, the pre-trained VGG16 is fine-tuned and retrained, and the retrained VGG16 is used to extract high-level implicit features of the 3D time-frequency spectrogram based on log-Mels, and the extracted high-level implicit features are re-described and mapped to bird species using the XGBoost model. Experimental results on BirdsData based on real natural scenes show the proposed method effectively break through the limitations of complex environments and achieve an accuracy of 95.13% on the test set.