多特征融合的鸟类物种识别方法

针对短时窗平均/长时窗平均算法从次声台站监测数据中提取的信号仍然包含噪声的问题,对支持向量机和人工神经网络的机器学习方法进行了研究。采用小波包分解的方法对信号进行重构,提取出各频带内的重构信号能量特征,对事件信号和噪声进行了识别实验,并分析了提高识别能力的方法,为工程应用提供理论参考。实验结果表明,在训练数据集不大的情况下,通过优化模型结构可以将两种方法的识别能力提高到可以接受的水平。

Bird species recognition method based on multi-feature fusion

The choice of input feature directly affects the classification performance of the deep learning, a multi-feature fusion recognition method was proposed to improve the classification performance of the bird species recognition model. In this method, firstly three kinds of spectrogram samples of vocalization signals were calculated through short time Fourier transform, Mel-frequency cepstrum transform and chirplet transform respectively, then three single feature models which based on VGG16 transfer learning were trained using these three kinds of spectrogram samples accordingly, modified weighted cross entropy function was used to fix the problem of imbalanced data set, the outputs of three models were fused to classify the spectrograms and realize the recognition of bird species. Taken the 35 kinds of bird in ICML4B database for study subject, the MAPs were compared, results show that the mean average precision (MAP) of feature fusion model is highest increased by 0.307 contrast to the single feature model; Three spectrogram durations, 100 ms, 300 ms and 500 ms were chosen to compare the test MAP of four models, the results reveal that the 300 ms duration is the best; the precision of 4 models with different SNR were compared, the precision reduction of feature fusion model as the SNR decreased is the least. The proposed model can achieve better performance with suitable duration, have anti-noise ability in some degree, and the trainable parameters are less, which is more suitable for birds with little samples.