应用无监督最大互信息算法分类鸟类叫声

Application of Unsupervised Maximum Mutual Information Algorithm to Classify Bird Songs

在建立鸟类叫声的分类模型时,由于自然界中具有准确标签的鸟类叫声数据较少,因此需要解决小样本下的模型训练问题。本文研究应用最大互信息的无监督网络来对鸟类叫声进行分类。通过同时提取梅尔图谱的高层语义特征和浅层特征并计算互信息,减少噪声特征的提取。训练时使用对抗样本,利用先验约束网络拉大不同类别之间的距离,降低模型对数据的依赖。实验证明,与现有无监督方法相比,利用最大互信息方法的无监督学习能够在鸟类叫声分类任务上取得最好的效果。

When establishing a classification model for bird calls,due to the limited number of accurately labeled bird call data in nature,it is necessary to solve the problem of model training in small samples.This article studies the application of unsupervised networks with maximum mutual information to classify bird calls.By simultaneously extracting high-level semantic features and shallow features of the Mel graph and calculating mutual information,the extraction of noise features is reduced.During training,adversarial samples are used,and a prior constraint network is used to widen the distance between different categories and reduce the model's dependence on data.Experiments have shown that compared to existing unsupervised methods,unsupervised learning using the maximum mutual information method can achieve the best results in bird call classification tasks.