Abstract:

[Objective] The accuracy of bird species classification can be improved by noise spectrogram filtering and removing from the sample set of original bird audio spectrograms. [Methods] Based on the convolutional neural network, this paper extracts the feature vector from the spectrogram, calculates the distance matrix of the feature vector with the Faiss algorithm library, and then uses the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm to filter out the noise spectrogram. Finally, the filtered spectrogram sample set is input into the classification model for bird species classification. [Results]Through this method, a large number of noise spectrograms are removed from the spectrogram sample set so that the accuracy of subsequent bird species classification has been improved. [Limitations] Because the clustering effect of the DBSCAN algorithm is greatly affected by the neighborhood threshold (ε) and density threshold (MinPts) parameters, we should explore adaptive methods to obtain parameter values in the future. [Conclusions] This paper combines the convolutional neural network and the density clustering algorithm in data mining and proposes a bird audio data preprocessing method for automatically noise spectrogram filtering, which provides a high-quality spectrogram sample set for subsequent bird species identification.

Key words: bird audio, spectrogram, data filtering, convolutional neural network, clustering