Multi-view features fusion for birdsong classification

• DOI: 10.1016/j.ecoinf.2022.101893
• Publication Date: 2022-11-03T17:07:34Z
• Authors: Shanshan Xie; Jing Lu; Jiang Liu; Yan Zhang; Danjv Lv; Xu Chen; Youjie Zhao
• Keywords: Mel-frequency cepstrum; Perceptron; Feature (linguistics)

As important members of the ecosystem, birds are good monitors ecological environment. Bird recognition, especially birdsong has attracted more and attention in field artificial intelligence. At present, traditional machine learning deep widely used recognition. Deep can not only classify recognize spectrums birdsong, but also be as a feature extractor. Machine is often to extracted handcrafted parameters. data samples classifier, directly determines performance classifier. Multi-view features from different methods extraction obtain perfect information birdsong. Therefore, aiming at enriching representational capacity single getting better way combine features, this paper proposes classification model based multi-view which combines by convolutional neural network (CNN) features. Firstly, four kinds extracted. Those wavelet transform (WT) spectrum, Hilbert-Huang (HHT) short-time Fourier (STFT) spectrum Mel-frequency cepstral coefficients (MFCC). Then CNN extract WT, HHT STFT minimal-redundancy-maximal-relevance (mRMR) select optimal Finally, three models (random forest, support vector multi-layer perceptron) built with probability results two types fused new Taking sixteen species research objects, experimental show that classifiers accuracy 95.49%, 96.25% 96.16% respectively for proposed method, than seven involved experiment. This method effectively perspectives signal. The comprehensively express bird audio itself, have higher lower dimension, improve classification.