AMResNet: An automatic recognition model of bird sounds in real environment. (December 2022)

• Record Type: Journal Article
• Authors: Xiao, Hanguang; Liu, Daidai; Chen, Kai; Zhu, Mi
• Is Part Of: Applied acoustics. Volume 201(2022)
• Publication Date: 2022-12
• Subjects: Bird sounds -- Automatic recognition model -- Residual network -- Deep Learning; Acoustical engineering -- Periodicals; Periodicals; 620.2
• DOI: 10.1016/j.apacoust.2022.109121
• Languages: English
• ISSNs: 0003-682X

Abstract

Highlights: An end-to-end method for automatic recognition of bird calls in real environments is proposed. A novel deep learning model, AMResNet, is established to automatically extract and select high-dimensional features to achieve a high classification accuracy. The combined feature sets as input give a more comprehensive representation of bird sounds and are more efficient. The method based on the combination of attentional mechanism and residual networks reduces the error of bird sounds recognition against other eight methods. Abstract: Birds are biological indicators reflecting environmental quality and its changes, so ecologists devote a great deal of attention to monitor their population trends. Automated acoustic recognition is regarded as an important technology to support bird monitoring and conservation activities. However, traditional bird sounds recognition methods have problems such as high cost, long time consuming and low precision. To address these problems, a novel automatic recognition model of bird sounds was proposed. The proposed model (AMResNet) was based on the combination of attentional mechanism and residual networks, which can automatically extract and select high-dimensional features to achieve a high classification accuracy. Attentional mechanism can improve the recognition efficiency by assigning appropriate weights to channels and space. Residual network can alleviate the problem of gradient disappearance and increase the information flow by skip connections. More importantly, an efficient combined feature was adopted in this paper, to give a more comprehensive representation of bird sounds. By trained and tested by 10-fold cross-validation in 12651 bird sound samples from the real environment, the AMResNet with combined features was proved to be appropriate for bird sound taxonomic problems, and dramatically outperform over other eight models, which contains two traditional models, a forward neural network, four main deep learning models and a latest vision transform model. The proposed model achieved a classification accuracy of 92.6%, which was 3.1% higher than the highest of other eight models, and it also achieved the best results in precision (97.6%), recall (97.3%) and F1-score (97.1%) across the 19 species.