Bird sound detection based on sub-band features and the perceptron model

Bird sound is a reliable indicator for monitoring bird population richness. In this paper, a bird sound detection method by using sub-band features and the perceptron is proposed. The Fisher Ratio (FR) is used for feature selection, and 23 dimensional features are extracted from the original 112 dimensional feature set. The feature dimensions of the selected feature set are about one-fifth of the original feature set, greatly reducing the computational complexity of the model. The accuracy, recall, precision, and F1-score achieved by perceptron model is 98.72 %, 98.90 %, 98.54 %, and 98.72 % respectively, and the prediction time is 40 ms. These results are also compared with those from a convolutional recurrent neural network (CRNN) and Random Forest (RF). Though the CRNN has similar accuracy, recall, precision and F1-score with the method proposed in this paper, its prediction time of 29 s is longer than the proposed model. The computational cost of RF is relatively acceptable, but its detection performance, except for recall rate, is significantly lower than the proposed model. The results show that the proposed method has excellent performance and low computational cost, and can be used for realtime bird sound detection.