基于信息熵与形态学的鸟鸣数据处理

Data processing of bird song based on information entropy and morphology

为了探索一种减少训练量并提高精度,且适应于卷积神经网络的预处理方法,以识别鸟鸣为例,基于信息熵以及形态学在图像处理上的应用,根据鸟鸣的间隔性特点,提出新的处理方案。利用形态学将音频片段归类为有效和噪声信号两类;使用加权的方法利用信息熵预测有效数据的分布。仿真结果表明,形态学使数据量减少且信息熵处理使单个数据稀疏化并起到滤波作用,在保持精度甚至精度提高的情况下缩短了训练时间,为卷积神经网络的数据预处理的简化提供了方向。

In order to explore a preprocessing method that reduces training cost,improves the accuracy and adapts to Convolutional Neural Network.Taking the identification of birdsongs as an example and according to the interval characteristics of the birdsongs,a new processing method is proposed based on the work of image processing with information entropy and morphology.Morphology is used to categorize the audio segment into two types:effective signals and noisy signals;Infonnation Entropy and weighted methods are used to predict the distribution of effective data.The simulation results show that the use of Morphology reduces the amount of training data and Information Entropy processing makes data sparse and can work as a filter,In general,the process shortens the training time while accuracy is even better,providing direction for the simplification of data preprocessing for Convolutional Neural Network.